271: Big Honkin' Die

00:00SPEAKER- Okay, Brad, I finally got rid of the last Dell 24.

00:04SPEAKERIt's either an '08 or '05 monitor in my house.

00:07SPEAKER- I'm less concerned about the '05 or the '08

00:10SPEAKERand more about the 24.

00:12SPEAKERYou had the 24 inch?

00:13SPEAKER- Yeah, so we got the 24s.

00:14SPEAKERWe all bought 24s.

00:15SPEAKERI don't remember why.

00:16SPEAKERI think we got, I think that we could buy them

00:18SPEAKERthrough the IPT department at our company.

00:20SPEAKER- Aha, that explains it. - And got a little bit

00:23SPEAKERof a deal, yeah.

00:24SPEAKER- Okay, we all bought the 20 inch version of that.

00:26SPEAKERWe're talking that mid 2000s Dell monitor

00:28SPEAKERthat literally everybody bought.

00:30SPEAKER- I think the 2405 was specifically the '05 model

00:33SPEAKERand the 2408 was specifically the '08 model.

00:35SPEAKER- That's correct.

00:36SPEAKER- We all bought the 20 inch 'cause even a 20 inch

00:39SPEAKERwas pretty expensive for an LCD monitor at that time.

00:41SPEAKERBut anyway.

00:43SPEAKER- I thought it was like 500 bucks, but anyway, yeah.

00:45SPEAKER- I think it was 600 for the 20 inch, if I remember.

00:47SPEAKER- They've been aging in my garage for a long time now.

00:50SPEAKERAnd when my daughter wanted a computer during the pandemic,

00:53SPEAKERrather than buy a monitor, which was kind of difficult

00:55SPEAKERto get as everybody's setting up their home offices,

00:57SPEAKERI just went out into the office, into the garage

00:59SPEAKERand grabbed the one that had the DisplayPort port

01:02SPEAKERand plugged it into her small form factor.

01:05SPEAKER- You went into your monitor seller,

01:07SPEAKERgrabbed one off the rack.

01:08SPEAKERThey've been aging.

01:09SPEAKERYou have some fine aged monitors here.

01:11SPEAKERThis is a nice 2005 vintage.

01:13SPEAKER- Let's see, I have a 3008 in there that's dead

01:17SPEAKERthat has some sort of bad capacitor

01:19SPEAKERthey should probably try to bring back to life.

01:20SPEAKER- Wait, how many Dell monitors do you have?

01:23SPEAKER- Oh, I, you know, there's four maybe.

01:25SPEAKER- Hang on, I took all that grief for having two Dell monitors

01:29SPEAKERfrom the era and all this time you had like three?

01:32SPEAKER- Yeah, but the difference is mine are just not in e-waste

01:36SPEAKERand yours you actually were using.

01:39SPEAKER- True, yes.

01:40SPEAKER- Yes, but yeah, so anyway,

01:43SPEAKERSanta brought her a new 24 inch,

01:48SPEAKER1080p, 144 Hertz gaming monitor.

01:51SPEAKER- Okay, my bad.

01:53SPEAKER- It's the zero bezel white frame.

01:56SPEAKERShe's very, she was very, the white,

01:57SPEAKERthe white was the important part it turned out.

01:59SPEAKER- I can say, man, I didn't know.

02:00SPEAKERI mean, I'm not surprised they make white monitors,

02:02SPEAKERbut I guess I have never seen one.

02:04SPEAKERDo they make other colors?

02:06SPEAKER- Mostly white.

02:06SPEAKERI think some people make pink ones probably

02:08SPEAKERif I had to guess, but it has so little frame,

02:11SPEAKERit doesn't matter.

02:12SPEAKERAnd since I put it on an arm,

02:13SPEAKERthen you can't see the base, which was white.

02:15SPEAKERSo it's kind of ridiculous that it's white, but.

02:18SPEAKER- Well, now you just need a white monitor arm.

02:20SPEAKER- I mean, I could turn the black one white

02:23SPEAKERwith a few minutes and spray paint.

02:25SPEAKER- Yeah, fair, fair.

02:27SPEAKER- What's the worst that happens?

02:28SPEAKERBut the thing about it was she wasn't excited about it.

02:31SPEAKERLike she wasn't excited about a monitor upgrade.

02:33SPEAKER- Hang on, was it a bigger monitor or was it the same size?

02:36SPEAKER- It's the same size, it's the same resolution.

02:37SPEAKERIt's thinner and faster.

02:39SPEAKERIt's 60 Hertz to, you know.

02:42SPEAKER- Yeah, that's something,

02:43SPEAKERthe high refresh is something you can't appreciate

02:45SPEAKERuntil you see it, to be fair.

02:47SPEAKER- Yeah, but when she sat down in front of her,

02:48SPEAKERshe's like, oh yeah, this is better, I like this.

02:51SPEAKER- Speaking of desk upgrades, I got a new plant.

02:53SPEAKER- Ooh, what's its name?

02:55SPEAKER- Around the holidays, I don't know if it has a name yet.

02:57SPEAKERIt was actually not a new plant.

02:59SPEAKERIt was a, we have a lot of plants in this house, by the way.

03:02SPEAKER- You do?

03:04SPEAKER- This was a hand-me-down from a different room.

03:06SPEAKERMy girlfriend has an extremely green thumb.

03:08SPEAKER- Okay, just like a cutting as a child of, or is this a?

03:12SPEAKER- It was a full potted.

03:13SPEAKER- Okay.

03:14SPEAKER- It's a ZZ plant.

03:15SPEAKER- What, it's ZZ?

03:17SPEAKER- ZZ, have you never heard of, ZZ stands for,

03:19SPEAKERit stands for something.

03:20SPEAKER- Oh, it's like ZZ top?

03:21SPEAKERDoes it have a beard?

03:22SPEAKER- Let's see, Zanzibar gem is one name.

03:25SPEAKER- Okay.

03:25SPEAKER- Zamioculcus is another name for it, I think.

03:29SPEAKERGot a lot of Z's in it.

03:30SPEAKER- Look, I was down at the dispensary the other day.

03:32SPEAKERI got some Zanzibar gem is all I'm saying.

03:34SPEAKERYou know what I mean?

03:35SPEAKER- Yeah, that's fine, I hear you.

03:36SPEAKERZamioculcus zamiopholia is the actual name

03:39SPEAKERof this plant apparently.

03:40SPEAKERIt's a thrives in low light plant,

03:42SPEAKERwhich was the key because there is low light in this office.

03:45SPEAKER- You know, it's funny, one of the projects

03:47SPEAKERthat I've had on my list for years

03:49SPEAKERand haven't gotten around to is,

03:50SPEAKERI would love to make the wall off to my,

03:53SPEAKERthe one that's off camera, the ones off to the side.

03:55SPEAKERI would love to just put a bazillion low light plants

03:59SPEAKERon there.

04:00SPEAKER- I love a wall garden.

04:02SPEAKER- Yeah, 'cause like this room is stuffy.

04:04SPEAKERThe windows are never open

04:06SPEAKERbecause it has to stay dark for lighting.

04:08SPEAKERAnd because I like the dark

04:09SPEAKERand I would love to have something living in here

04:14SPEAKERthat's not me and whatever food I drop on the ground.

04:17SPEAKER- Yeah, I can't say that the number of plants in here

04:21SPEAKERis affecting the oxygen content of our apartment,

04:23SPEAKERbut I can't say that it's not.

04:25SPEAKER- Look, it takes a lot of plants

04:27SPEAKERto affect the oxygen content of your apartment

04:29SPEAKERand just one person to bring it down.

04:32SPEAKER- Well, hey, look, there are other benefits.

04:34SPEAKERSometimes you sit in a largely lightless room all day.

04:38SPEAKERSometimes it's nice to just stare at something green.

04:40SPEAKER(upbeat music)

04:47SPEAKER(upbeat music continues)

05:08SPEAKER- Welcome to Brad and Will Made a Tech Pod.

05:11SPEAKERI'm Will.

05:12SPEAKER- I'm Brad.

05:13SPEAKER- Brad, it's the most wonderful time of year.

05:15SPEAKER- What new video card time?

05:17SPEAKER- New video card season.

05:19SPEAKER- I have to confess,

05:20SPEAKERI'm having a little bit of a hard time

05:21SPEAKERtalking about new video cards right now.

05:22SPEAKERI'm on such pins and needles over the intermittent issues

05:26SPEAKERwith my audio set up here.

05:27SPEAKER- Oh no, no.

05:29SPEAKER- Now I'm blowing up the,

05:32SPEAKERall of my editing around the problems

05:34SPEAKERby just talking about it.

05:35SPEAKERI keep, I keep fuzzing out, you know,

05:37SPEAKERmaybe we should leave one in the show

05:38SPEAKERand people listening who know audio can be like,

05:40SPEAKER"Oh, that's that kind of interference."

05:42SPEAKER- Just put it in at the end.

05:44SPEAKERDrop it in all the way at the end

05:45SPEAKERand let people get in there.

05:46SPEAKER'Cause it sounds kind of, it's kind of herdy

05:48SPEAKERif you have your headphones on is what I would say so.

05:51SPEAKER- What I'm talking about is since I got this channel strip

05:53SPEAKERand I've, you know,

05:54SPEAKERI've got a fairly complex analog audio mic situation

05:57SPEAKERgoing on here and I'm just picking up

05:59SPEAKERsome weird interference every so often.

06:01SPEAKERLargely when we record this podcast,

06:03SPEAKERit doesn't happen in like three weeks

06:04SPEAKERand then this morning it's been happening again pretty badly

06:07SPEAKERand it makes it hard to record

06:08SPEAKERwhen you just randomly fuzz out like crazy.

06:11SPEAKER- It's not even like a fuzz out, it's like.

06:14SPEAKER- It's kind of scratchy.

06:15SPEAKER- Yeah, you sound, yeah, it sounds like,

06:18SPEAKERit sounds like you're playing an audio tape

06:19SPEAKERthat has some schmutz on the tape

06:21SPEAKERand it's messing with the read heads.

06:23SPEAKERYou got stuff on your read heads?

06:24SPEAKER- No schmutz, please.

06:26SPEAKER- No schmutz.

06:27SPEAKERWell, while you've been dealing with your audio issues,

06:32SPEAKERI have been in video card testing times.

06:37SPEAKERNvidia, you and I talked on a PC World live stream

06:39SPEAKERa couple weeks ago during CES about the Nvidia keynote.

06:44SPEAKERWe talked over the Nvidia keynote.

06:45SPEAKER- True.

06:46SPEAKER- Where they announced these,

06:47SPEAKERbut I've had hardware since Monday, so about five days now.

06:52SPEAKER- Oh.

06:53SPEAKER- Only the 1590, they only sent me a 1590.

06:57SPEAKER- Yeah, they're being kind of weirdly cagey

06:59SPEAKERor maybe not weirdly,

07:00SPEAKERbut they're being kind of cagey about the 5080, right?

07:02SPEAKERLike it seems like that one might be the least impressive

07:05SPEAKERrelatively by price of all of them.

07:08SPEAKER- I mean, they always do a staggered rollout on these

07:10SPEAKERjust so they don't have four, they announced four cards,

07:14SPEAKERa 1590, a 5080, a 5070 Ti, and a 5070.

07:18SPEAKERThey're making founders additions for all of those

07:20SPEAKERexcept for the 5070 Ti.

07:22SPEAKER- Wait, really?

07:23SPEAKER- Yeah, I think- - Oh, that sucks.

07:24SPEAKER- It seems like 5070 Ti might be

07:26SPEAKERthe price to performance winner.

07:28SPEAKER- I think that that's probably the concession

07:30SPEAKERto the ad board vendors is that they're not gonna compete

07:32SPEAKERwith them on the one that's gonna be the sweet spot.

07:35SPEAKER- That's the one everybody's gonna want.

07:36SPEAKERSo here's, we'll throw you a bone

07:38SPEAKERbecause we make all our money from data centers.

07:40SPEAKER- That would be my guess.

07:42SPEAKER- We'll let you sell all of the ones

07:44SPEAKERthat people actually want.

07:45SPEAKER- Yeah, so it's the first consumer pass

07:49SPEAKERon their new architecture.

07:50SPEAKERBlackwell has been out for a while now for data center users

07:54SPEAKERand for like heavy industrial computer users.

07:58SPEAKERBut this is the first upgrade from Lovelace,

08:01SPEAKERfrom Ada on the consumer side.

08:06SPEAKERSo 4090s came out two years ago,

08:08SPEAKERa little bit under two years ago, I guess now.

08:10SPEAKER- Yes, that sounds right. - Is that right?

08:11SPEAKER- I think that's right.

08:13SPEAKER- And then the other cards in that family followed.

08:17SPEAKERAnd that's the way this is gonna roll out.

08:18SPEAKERSo the embargo for the 5090s was Thursday

08:22SPEAKERand the cards are gonna be available on the 30th, I think,

08:25SPEAKERin I'm gonna guess extremely limited quantities.

08:30SPEAKERBut the other cards are gonna follow shortly thereafter.

08:34SPEAKERAnd I'm sure people are already testing 5080s

08:38SPEAKERand so on down the line.

08:40SPEAKER- Yeah, will the quantities be that limited?

08:42SPEAKERI mean, I'm sure there'll be somewhat,

08:43SPEAKERbut like, I mean, we're not in the like early pandemic,

08:46SPEAKERlike everybody's fighting to get a 3080 type territory,

08:49SPEAKERyou know, like the PS5 Pro came out

08:50SPEAKERand there were just plenty to go around a couple months ago.

08:53SPEAKERLike it feels like things are better.

08:55SPEAKER- I think it's gonna be a demand situation

08:58SPEAKERmore than a supply situation.

09:00SPEAKER- That seems crazy to me

09:01SPEAKERbased on the relative performance increase of these things

09:05SPEAKERcompared to the price and the heat and everything,

09:06SPEAKERbut go on.

09:07SPEAKER- So the reason I say that is it's on the same process node,

09:10SPEAKERTSMC process node as the 4090, right?

09:13SPEAKERSo this is, there's no new fab technology here

09:17SPEAKERthat's gonna make these things hard to produce.

09:19SPEAKERAnd in fairness, the 4090 supply has dried up

09:22SPEAKERover the last couple of months.

09:23SPEAKERSo that implies pretty strongly that they've shifted.

09:27SPEAKERYou can infer when they shifted over

09:29SPEAKERfrom 4090 production to 5090 production.

09:32SPEAKER- And also, as I understand,

09:33SPEAKERor at least looking at the sort of block diagram,

09:35SPEAKERthis is a big honking dice.

09:38SPEAKERThey get fewer per wafer, I assume.

09:40SPEAKER- It's a big, big honking die.

09:44SPEAKERSo, I mean, I guess let's start by talking about

09:47SPEAKERthe hardware and what they changed on the architecture

09:49SPEAKER'cause there's some big stuff that's changed.

09:51SPEAKERThe main thing is that it has just a whole buttload

09:56SPEAKERmore compute across the board.

09:59SPEAKERThe memory bandwidth has increased from the 4090 as well.

10:02SPEAKERIt's almost a 2X, it's like 1.7 gigabytes per second

10:06SPEAKERversus 1,000, sorry, 1,792 gigabytes a second

10:10SPEAKERversus 1,008 gigabytes a second.

10:13SPEAKERAnd they do that by using GDDR7 and a wider memory bus.

10:17SPEAKERYeah, the memory bus jumps from 512 bits,

10:20SPEAKERfrom 384 bits to 512 bits on the 4090 to the 5080.

10:24SPEAKERThe reason the memory bandwidth increase is important

10:26SPEAKERis twofold.

10:26SPEAKEROne, for the GPU compute people, it's important.

10:29SPEAKERBut for people who play video games,

10:31SPEAKERthere's a direct correlation between the number of pixels

10:34SPEAKERyou can render traditionally using raster pipelines

10:36SPEAKERand the memory bandwidth usually.

10:39SPEAKERI mean that you can hit bottlenecks in other places,

10:41SPEAKERbut for the most part, you're limited by,

10:44SPEAKERwhen memory bandwidth goes up, frames go up.

10:47SPEAKERSo they redesigned the ray tracing cores.

10:51SPEAKERSo they are more efficient and faster and more capable.

10:55SPEAKERSo that means you can do more rays on each frame

10:58SPEAKERthan you could previously, which means less noise,

11:01SPEAKERless kind of artifacting, less other shenanigans

11:04SPEAKERthey have to do to make the ray fill-in look good.

11:07SPEAKER- Does that mean if you're running a ray tracing game

11:09SPEAKERon a 5090, it is actually casting more rays

11:11SPEAKERthan it would on a 4090?

11:13SPEAKER- Yeah, that's my understanding.

11:14SPEAKER- It's not just that you're getting more frames

11:16SPEAKERwith the same number of rays cast?

11:17SPEAKER- Yeah, and the same thing on lower end cards too.

11:21SPEAKERSo the 4070, 4080, 5080, 5070 will do fewer,

11:26SPEAKERlike you're doing more work with fewer rays

11:29SPEAKERon the lower end cards.

11:30SPEAKERSo there should be an IQ improvement.

11:32SPEAKER- Yeah, this is very in the weeds.

11:34SPEAKERThen we probably need to talk

11:35SPEAKERto a graphics engineer to know this.

11:37SPEAKERI wonder how much of that is automated,

11:38SPEAKERhow much of that is in drivers or the APIs,

11:40SPEAKERor do you have to manually tune performance

11:43SPEAKERfor how many rays you cast based on hardware?

11:45SPEAKEROr does that just scale automatically based on,

11:48SPEAKERlike I said, is that handled at the library level?

11:52SPEAKER- I wonder if it's a choice.

11:54SPEAKERI wonder if you make the choice

11:55SPEAKERthat you wanna have X number of rays,

11:57SPEAKERand then if it, I actually don't know how that works.

12:00SPEAKERThat's really, that's a good question.

12:01SPEAKER- Although, honestly, the denoising has gotten good enough

12:04SPEAKERthat it, is it actually, I mean, obviously it's better

12:07SPEAKERto have real data sampled than interpolated data,

12:10SPEAKERbut it's still, I don't know if that's a huge advantage,

12:14SPEAKERbecause even the old cards look quite good

12:16SPEAKERwith the current denoising and ray reconstruction

12:19SPEAKERand so forth.

12:20SPEAKER- Well, I was gonna say, yeah, the ray reconstruction,

12:22SPEAKERwhich uses AI to basically fake more rays,

12:24SPEAKERit turns out, is quite good.

12:27SPEAKERAnd is that and the denoising make the ray,

12:31SPEAKERlike, it's, the ray tracing is fine on even,

12:36SPEAKERlike, lower end cards that don't do that these days.

12:40SPEAKERWe've gotten better at ray tracing.

12:41SPEAKER- Right, like even, I mean, we're really in the weeds here,

12:44SPEAKERbut even on like a launch 3080 that I had,

12:46SPEAKERI didn't really, I didn't really have complaints

12:48SPEAKERabout like the image quality of the ray,

12:49SPEAKERI mean, obviously it's great for it to get better over time,

12:51SPEAKERbut that was pretty good, even from the jump, I thought.

12:55SPEAKER- Well, I mean, the big difference is that you can do,

13:00SPEAKERlike, you can do more complicated ray tracing games.

13:03SPEAKERLike, if you look at the lighting model

13:04SPEAKERin Indiana Jones, right, it's, they've done a full,

13:09SPEAKERit's the closest we have

13:10SPEAKERto a full ray tracing implementation.

13:13SPEAKER- Oh, even more so than Cyberpunk?

13:16SPEAKER- Since they added path tracing to Cyberpunk,

13:18SPEAKERlike, those are the two, those are the two high bars

13:21SPEAKERas far as I can tell, in terms of what's been,

13:24SPEAKERwhat's being used, what makes the card do the work,

13:27SPEAKERand what looks good, so.

13:29SPEAKER- I've got a fire up in Indiana Jones.

13:30SPEAKERI've got a 4090, which is not the fancy new hotness,

13:34SPEAKERbut still, I need to see that path tracing implementation.

13:38SPEAKER- It is, I'm gonna go and tell you,

13:39SPEAKERit's nice to have a game that is both gorgeous and fun

13:43SPEAKERto experience this stuff with.

13:45SPEAKER- Sure.

13:46SPEAKER- I've been blown away with it.

13:48SPEAKER- About a stealth diss on Cyberpunk?

13:51SPEAKER- I find Cyberpunk to be almost incomprehensible.

13:53SPEAKER- Yeah, there's a lot going on there, sure.

13:55SPEAKER- Yeah.

13:57SPEAKEROkay, so, four SKUs, 5090, it's $2,000, 32 gigs of RAM,

14:04SPEAKERhuge memory bandwidth increase over last gen and the 5080.

14:09SPEAKERThe 5080 is a grand, it's 16 gigs of RAM,

14:12SPEAKERthe 5070 Ti is $750 and it's 16 gigs of RAM,

14:17SPEAKERand the 5070 Vanilla is 550 bucks and 12 gigs of RAM.

14:22SPEAKERSo, I don't have any experience hands-on

14:24SPEAKERwith the 5070, 5070 Ti, or 5080,

14:27SPEAKERbut looking at the memory configurations,

14:29SPEAKERI assume that you're looking at 5070 and 5070 Ti

14:33SPEAKERas 1440p cards, and 5080 as a 4K card,

14:39SPEAKERand 5090 for people who buy the fastest card

14:42SPEAKERor do GPU computer stuff.

14:43SPEAKER- Yeah.

14:45SPEAKER- The memory bandwidth on the 5080,

14:48SPEAKERjust so you have a point of comparison,

14:50SPEAKERit's double, it's half of what the 5090 is.

14:54SPEAKERSo, it's 1792 gigabytes per second on the 5090

14:58SPEAKERand 960 gigabytes on the 5080,

15:02SPEAKERand then it goes down further from there

15:04SPEAKERto the 5070 Ti and the 5070 at 896 and 672 respectively.

15:11SPEAKER- Well, I mean, I started to say the 5080

15:13SPEAKERdoesn't look like a great deal, but also 5080

15:16SPEAKERis half the price of the 5090

15:17SPEAKERand also half the memory bandwidth.

15:19SPEAKERSo, I guess that, and I think it's roughly

15:21SPEAKERhalf the compute units as well, right?

15:23SPEAKERSo, I guess it is actually that the price does scale

15:26SPEAKERto the capabilities pretty directly.

15:28SPEAKER- Yeah, I think my guess is that when we get testing

15:32SPEAKERand benchmarks on the 5080s, which I think is,

15:35SPEAKERI don't know if I'm allowed to say when the bar goes up,

15:37SPEAKERbut I don't have any firsthand knowledge of this.

15:43SPEAKERI haven't been involved in the testing on that card at all

15:44SPEAKER'cause I've been working on DLSS stuff

15:46SPEAKERsince the 5090 reviews were done.

15:49SPEAKERIt seems like, especially if you play at 4K,

15:54SPEAKERit seems like it's gonna be a sweet spot for a grand.

15:56SPEAKERIt seems like they were aware of the problems with the 4080,

16:00SPEAKERwhich they kind of underperformed for the price last gen.

16:04SPEAKERAnd I am optimistic about this, but like I said,

16:07SPEAKERI don't have any inside knowledge

16:08SPEAKERand I'll find out when everybody else does.

16:11SPEAKER- All right.

16:12SPEAKERThe big thing for the 5080 and 5090 though,

16:17SPEAKERis the new cooler.

16:19SPEAKERSo, 4090s, if you spent some time with 4090s,

16:22SPEAKERfor the most part, they started as big honking cards

16:25SPEAKERand went up from there.

16:26SPEAKER- Mine is a Zotac, my 4090.

16:28SPEAKER- Yeah.

16:29SPEAKER- And to be fair, I pulled the stock fan shroud off

16:31SPEAKERand got to put a custom one on

16:33SPEAKERso I could put Noctua fans on that thing.

16:35SPEAKER- Oh my God.

16:36SPEAKER- So, we would be extremely silent.

16:38SPEAKERThe fans that ship with that thing sounded like terrible.

16:40SPEAKER- Yeah, it's a loud boy.

16:43SPEAKER- And now it's completely silent.

16:44SPEAKERAnyway, like the point is I made my card slightly bigger

16:46SPEAKERthan it already was, but it was already very big.

16:49SPEAKERLike the launch 4090s were gigantic.

16:51SPEAKER- Yeah, the Founders Edition was a three slot card

16:55SPEAKERthat was also very, very long

16:57SPEAKERand didn't fit in a lot of cases.

16:59SPEAKERThe new 5090, 5080 design is a two slot.

17:03SPEAKERThey call it a double blow through.

17:06SPEAKERPreviously on the 4090, they did a cutout.

17:09SPEAKERThey did like a diagonal cutout out of the end of the board

17:12SPEAKERso that the second fan would blow through

17:14SPEAKERand you'd get straight through airflow.

17:16SPEAKERThe thing that they realized is that when you have to turn,

17:19SPEAKERwhen the air has to turn from the intake

17:21SPEAKERto the exhaust on the card,

17:23SPEAKERlike it blows in and hits a radiator

17:25SPEAKERand then has to turn 90 degrees

17:26SPEAKERand go out either the back of the case

17:29SPEAKERor out the front of the card,

17:31SPEAKERthat reduces efficiency dramatically.

17:34SPEAKERSo, by doing these cutouts,

17:36SPEAKERthey were able to increase the performance of the cooler.

17:40SPEAKERAnd on this, on the 55 series,

17:42SPEAKERthey went from having one cutout

17:43SPEAKERto just having the board floating in the middle of the card

17:48SPEAKERwith some daughterboards

17:49SPEAKERthat go to the exterior monitor connectors

17:53SPEAKERand the PCI express slot.

17:55SPEAKERAnd then the rest of it is just cooling.

17:57SPEAKERSo, it's huge radiators,

17:59SPEAKERthere's a vapor chamber over the die and the RAM,

18:02SPEAKERand then there's heat pipes

18:03SPEAKERthat carry the evaporated liquid from the vapor chamber,

18:07SPEAKERthe gas, the vapor, into the cool areas,

18:12SPEAKERcools them down and then wicks the liquid

18:13SPEAKERback into the center so it can evaporate again.

18:16SPEAKER- That sounds complex.

18:17SPEAKERI saw, I wish I could remember who this was on Blue Sky

18:20SPEAKERthat was talking specifically about the daughterboard

18:23SPEAKERor one of the daughterboards,

18:24SPEAKERbecause if you're going to separate the functionality

18:27SPEAKERof something that's moving this much data this fast

18:29SPEAKERinto multiple boards,

18:31SPEAKERthen I think they have to be incredibly electrically sound,

18:33SPEAKERright, when they were talking about the nature of the build

18:36SPEAKERof the daughterboard and the interconnect between it

18:38SPEAKERand the main board is, from an electrical standpoint,

18:41SPEAKERis wildly impressive because they kind of have to be.

18:44SPEAKER- Yeah, 'cause you think about making a PCI express lane

18:48SPEAKERaway from the motherboard is something we've been doing

18:50SPEAKERfor years now with vertical GPU mounts, right?

18:54SPEAKERSo, you basically plug a card in

18:55SPEAKERand then you have this big, giant, wide, firm, stiff cable

19:00SPEAKERthat goes into a female PCI express slot

19:02SPEAKERand then you plug that into the bottom of your video card

19:04SPEAKERand it just kind of hangs there, right?

19:06SPEAKERThe thing that they have going from the PCI express mount

19:10SPEAKERon the back of the main board of this card

19:12SPEAKERto the daughter card is like a tiny,

19:14SPEAKERI don't actually think they showed pictures of it,

19:16SPEAKERbut the connector is very small.

19:18SPEAKERIt's like a ZIF connector

19:19SPEAKERand the fact that they have the whole thing running there

19:22SPEAKERis, I think, is pretty impressive.

19:23SPEAKERGetting the connectors for the displays is less so,

19:25SPEAKERI would say, 'cause that's pretty,

19:28SPEAKERthose are solved problems.

19:29SPEAKERThose are just frames at that point.

19:32SPEAKERThat's not that much data.

19:33SPEAKERThe real question, though,

19:34SPEAKERis how well does this actually cool?

19:36SPEAKERI feel like I've seen mixed results

19:39SPEAKERfrom the initial testing.

19:40SPEAKER- It runs a little hotter than a 4090 typically under load.

19:44SPEAKERWhen I was doing testing of Indiana Jones, stuff like that,

19:48SPEAKERI was seeing 63-ish,

19:51SPEAKERthe NVIDIA card was reporting 63 Celsius temperatures,

19:54SPEAKERwhich is not, it's hot for a 4090,

19:58SPEAKERit's not hot for a 3080 or a 2080.

20:02SPEAKER- Like that's well within spec, right?

20:04SPEAKER- Yeah, and versus 57C for the 4090

20:11SPEAKERat the same kind of load.

20:12SPEAKER- Okay, that's not a huge disparity.

20:14SPEAKER- It is not.

20:15SPEAKERNow, the thing I will say is when I play a game

20:17SPEAKERthat uses full ray tracing at 4K,

20:20SPEAKERit makes, it generates a lot of heat.

20:23SPEAKERLike you can tell it's actually working.

20:25SPEAKERI often, I mostly play at 1080p

20:28SPEAKER'cause my gaming monitor is a 1080p 360 monitor.

20:33SPEAKERBut I did have to change the airflow

20:36SPEAKERin my case a little bit.

20:37SPEAKERSo I, my top back, so the way the air works on this

20:41SPEAKERis that the fans are on the bottom.

20:42SPEAKERIf you're in a tower, the fans are facing down,

20:45SPEAKERthe fans push air up through the radiators

20:48SPEAKERand then it comes out the top.

20:49SPEAKERThere's a little bit of a,

20:51SPEAKERSteve over at Gamers Nexus did a good video

20:53SPEAKERillustrating this with his weird camera setup

20:57SPEAKERthat shows air turbulence.

20:59SPEAKERAnd it has two holes that come out of the top of the card

21:04SPEAKERthat angle the air at like a 35 degree angle up

21:09SPEAKERso that it, so that it separates the hot air on the top

21:13SPEAKERfrom the cool air on the bottom of the case.

21:15SPEAKERAnd as a result, you want more exhaust on the top

21:19SPEAKERthan you probably have right now,

21:21SPEAKERespecially if you have fans blowing air in on the top,

21:24SPEAKERyou might wanna actually flip some of those around.

21:26SPEAKERI just turned up the exhaust fan a fair amount

21:29SPEAKERand that seemed to do the job.

21:32SPEAKERBut you're gonna, if you're playing games at 4K

21:34SPEAKERon this thing, especially the two slot designs,

21:37SPEAKERI think you're gonna wanna pay attention to cooling

21:39SPEAKERmore than you have in the past, especially with the 4090.

21:42SPEAKERWhich just to be clear, 4090 had one blow through

21:44SPEAKERand then one that exhausted, which did good work.

21:49SPEAKERSo I think the board design's pretty nifty.

21:52SPEAKERLike I don't think the two slot design matters

21:55SPEAKERfor gamers as much.

21:57SPEAKERI think it actually is a kind of a selling point

21:59SPEAKERfor the AIB, the aftermarket, not aftermarket,

22:02SPEAKERbut the cards made by people that aren't Nvidia

22:06SPEAKERbecause they'll have bigger, more traditional coolers

22:08SPEAKERthat maybe run a little hotter

22:11SPEAKERor maybe run a little louder,

22:13SPEAKERbut are more effective at exhausting heat from the system.

22:16SPEAKER- Yeah, and before we move on too far,

22:18SPEAKERI mean, we hinted at this earlier,

22:19SPEAKERbut not to bury the lead.

22:21SPEAKERLike the GPU is way bigger because like you said,

22:24SPEAKERit's on the same process.

22:25SPEAKERThey have achieved better performance

22:26SPEAKERby making a bigger chip than last time, right?

22:28SPEAKERLike we're not in an era of increasing efficiency

22:31SPEAKERleading to increased performance.

22:32SPEAKERLike it is simply a bigger part is a large part

22:36SPEAKERof how they're getting better performance,

22:38SPEAKERwhich is what has led

22:39SPEAKERto the higher power consumption and heat.

22:41SPEAKER- Yeah, that's not necessarily uncommon for GPUs though.

22:44SPEAKEROften they'll stay on process sizes

22:49SPEAKERfor longer than your typical CPU will, right?

22:52SPEAKER- Yeah, sure.

22:52SPEAKER- Because unlike CPUs, which you see an upside,

22:57SPEAKERa maximum upside to multi-threading,

22:59SPEAKERwith GPUs adding wider buses just means more performance

23:04SPEAKERand has since the dawn of time.

23:06SPEAKER- Right, but it is atypical for Nvidia though,

23:09SPEAKERbecause going from the 20 to the 30

23:11SPEAKERand then the 30 to the 40,

23:12SPEAKERthere were pretty big efficiency increases

23:14SPEAKERin both those cases.

23:14SPEAKERSo people have gotten a little used to

23:17SPEAKEReach Nvidia generation bringing

23:19SPEAKERmassive efficiency improvements,

23:20SPEAKERand that's not the case here.

23:22SPEAKER- Yeah, so they're both on TSMC five nanometer

23:24SPEAKERand the 50 90, the GB 202 is 750 millimeters,

23:30SPEAKERsquare millimeters, and the 40 90,

23:33SPEAKERthe AD 102 is 609 square millimeters.

23:36SPEAKERSo it's a little bit bigger.

23:37SPEAKERIt's not a lot bigger,

23:39SPEAKERbut it's not an insubstantial amount.

23:41SPEAKER- Yeah, it's like a 25% increase.

23:43SPEAKER- Yeah, that seems about right.

23:45SPEAKERI mean, the nice thing about them staying

23:47SPEAKERon the same process for this

23:48SPEAKERis it does make the possibility for a TI die shrunk version

23:52SPEAKERdown to one of the other nodes that,

23:54SPEAKER'cause TSMC four and some other ones,

23:56SPEAKERoften on GPUs, they stay on the bigger nodes

23:58SPEAKER'cause they'd rather have supply

24:00SPEAKERthan the heat matters less than having ready supply.

24:05SPEAKER- I guess all of that actually does kind of lead

24:07SPEAKERto the next big topic,

24:09SPEAKERwhich is all the AI stuff for, et cetera, et cetera,

24:12SPEAKER'cause like that's kind of where they're seeing

24:15SPEAKERa lot of the big performance improvements on new features

24:17SPEAKERis in the machine learning based stuff,

24:20SPEAKERnot in raw performance.

24:22SPEAKER- Well, so the raw performance looking at the numbers

24:26SPEAKERwas depending on the benchmark was between 30

24:29SPEAKERand 10 or 12%, up to 45 in some cases

24:34SPEAKERon some weird outliers where this happened

24:36SPEAKERto hit bottlenecks particularly well,

24:38SPEAKERbut call of duty on, I'm using PC world's benchmarks

24:42SPEAKERjust 'cause that's what I'm most familiar with,

24:46SPEAKERsaw like a 12% lift on call of duty, which feels low.

24:50SPEAKER- That's just for straight raster increase, to be clear.

24:53SPEAKER- That was for straight raster increase.

24:55SPEAKERYeah, so I guess it's the way to look at this

24:57SPEAKERis there's three different kind of categories

25:00SPEAKERof performance testing here, right?

25:03SPEAKEROne is the straight raster, which is no ray tracing.

25:07SPEAKERIt's just rendering frames with light maps

25:09SPEAKERand the way we've been doing stuff for 20 years.

25:11SPEAKERThat's in recent generations,

25:14SPEAKERthat's where we've seen the least lift generally.

25:16SPEAKERThat's also often tied to memory bandwidth increase,

25:18SPEAKERwhich is why I think we see the lift that we do here.

25:23SPEAKERFor ray tracing performance,

25:26SPEAKERray tracing performance is usually gated

25:28SPEAKERby the number of rays that like the speed

25:30SPEAKERat which you can generate the rays

25:31SPEAKERthat are needed to render the scene.

25:33SPEAKERAnd in this case, pretty good, number's pretty good.

25:37SPEAKERThe new ray tracing cores in Blackwell had significant

25:43SPEAKERperformance improvements kind of across the board.

25:46SPEAKERI'm looking for my percent increase chart here.

25:49SPEAKERAnd yeah, Cyberpunk with FrameGen off saw a 30% lift

25:56SPEAKERfrom 50/90 to 40/90, which feels pretty good to me.

26:00SPEAKER- Yeah, I feel like ray tracing improvements

26:03SPEAKERare still somewhat academic.

26:04SPEAKERI mean, obviously people do play half-race games now,

26:07SPEAKERso it does matter, but it's still, to me,

26:09SPEAKERthat's still largely a gauge of, okay,

26:12SPEAKERhow fast are we getting to the all ray tracing future?

26:15SPEAKER- Well, and there's like five path trace,

26:18SPEAKERthere's not a ton of path traced games, right?

26:20SPEAKER- Yeah, that's what I mean.

26:21SPEAKERI mean, that's really the big question

26:23SPEAKERwith ray tracing continues to be,

26:24SPEAKERwhen do we see the first ray tracing game

26:26SPEAKERthat only ships with a ray traced renderer?

26:28SPEAKERI'm still guessing maybe another 10 years from now.

26:32SPEAKER- It'll be either a PC exclusive title

26:34SPEAKERor it'll be next generation consoles.

26:36SPEAKER- That's the thing, it has to.

26:37SPEAKER- It has to descend from the consoles

26:39SPEAKERhaving the grunts to also do that,

26:40SPEAKER'cause they're never gonna do that just for a PC only game.

26:43SPEAKER- Yeah, that's exactly right.

26:45SPEAKERLike there are a handful of titles

26:47SPEAKERthat have real modern lighting models available

26:50SPEAKERif you have the right hardware for it,

26:51SPEAKERlike Star Wars Outlaws, Black Myth, Wukong, Indiana Jones,

26:56SPEAKERlike it's happening more.

26:58SPEAKERFor a long time, it was just cyberpunk,

27:00SPEAKERand now we have other stuff coming, which is exciting.

27:03SPEAKER- And cyberpunk was, the cyberpunk implementation

27:05SPEAKERwas like practically a research project

27:07SPEAKERin terms of the way Nvidia talked about it,

27:08SPEAKERbut now it's just kind of showing up

27:10SPEAKERin games slowly here and there.

27:11SPEAKER- Yeah, I think that's a fair way to say it.

27:16SPEAKERYeah, so like the uplift is pretty good

27:19SPEAKERon both raster and ray tracing.

27:21SPEAKERLike I would say more than acceptable for gen over gen

27:25SPEAKERwhen you add in the multi-frame gen stuff

27:27SPEAKERthat they've added with DLSS for.

27:29SPEAKERSo DLSS started out as super,

27:32SPEAKERlike a machine-augmented super,

27:34SPEAKERmachine learning's augmented supersampling

27:36SPEAKERto take a low resolution frame and make it big.

27:41SPEAKER- Which works very well.

27:42SPEAKER- Yeah, it's great.

27:43SPEAKER- Like it's proven, it feels like proven technology,

27:45SPEAKERlike they've ironed out so many of the kinks

27:46SPEAKERthat depending on who you ask,

27:48SPEAKERI think some people would say

27:49SPEAKERthat the machine learning upscaled frames

27:51SPEAKERare better than the raw frames at this point,

27:53SPEAKERor at least on par.

27:54SPEAKER- So when it works right, absolutely,

27:57SPEAKERI would say that's absolutely the case.

27:59SPEAKERI've been doing some direct comparisons

28:01SPEAKERbetween raw and different versions of DLSS,

28:05SPEAKERand more on that soon.

28:08SPEAKERI've not quite reached full conclusions yet,

28:11SPEAKERbut it's a little messier than I expected

28:13SPEAKERwhen you dig into the frames.

28:14SPEAKERBut the point is these cards are running

28:16SPEAKERat 150 or 200 frames a second,

28:18SPEAKERand at that speed, you don't notice the mistakes, right?

28:22SPEAKERSo DLSS used to mean upscaling.

28:24SPEAKERNow it's the name for the family

28:26SPEAKERof all of the kind of AI acceleration stuff.

28:31SPEAKERThat started, it seems like, last generation,

28:33SPEAKERand I think I didn't clock it then,

28:34SPEAKERbut yeah, so the DLSS super sampling, upscaling,

28:41SPEAKERthe big thing is that they have added a new model.

28:43SPEAKERSo previously, it used a convolutional neural network,

28:46SPEAKERwhich they've kind of been around for a long time.

28:50SPEAKERThey work by comparing adjacent pixels

28:53SPEAKERand saying, oh, if this pixel's black,

28:55SPEAKERthen usually, if this pixel's black

28:56SPEAKERand the pixel next to it is pink,

28:58SPEAKERthen usually the next pixel's gonna be blue, right?

29:01SPEAKERAnd they do that on a kind of very regional basis,

29:04SPEAKERwhich is why often if you had things that looked like hands,

29:07SPEAKERit would be like, oh yeah, this is flesh tone,

29:11SPEAKERshadow, flesh tone, shadow.

29:13SPEAKERThat's exactly right.

29:14SPEAKERThat's why we have six fingers

29:15SPEAKERon some of these frames sometimes, right?

29:17SPEAKERIt's 'cause it just keeps putting flesh tones and shadows

29:19SPEAKERand it doesn't know the difference.

29:21SPEAKERThey've switched to a transformer model,

29:23SPEAKERwhich is what Chad GPT and a lot of the,

29:25SPEAKERI don't want to say smarter,

29:27SPEAKERbecause that's not exactly right,

29:29SPEAKERbut the more aware of context,

29:32SPEAKERit lets the model be more aware of context.

29:35SPEAKERSo it knows if this is a screen full of dogs,

29:39SPEAKERthere's likely to be other things that look like dogs

29:41SPEAKERand it'll fill in things that it knows

29:43SPEAKERto be associated with dogs.

29:48SPEAKERIt's basically, there's a bunch of terrible analogies

29:51SPEAKERfor this, but the idea, with these,

29:54SPEAKERyou're always just going through,

29:56SPEAKERplacing vectors through arrays,

29:59SPEAKERtwo-dimensional arrays of common,

30:02SPEAKERof whatever you're looking at, pixels, words, whatever.

30:05SPEAKERIn this case, you're looking at,

30:06SPEAKERinstead of going at a two-dimensional array

30:08SPEAKERwith the transformer, you're adding more dimensions.

30:10SPEAKERSo you can kind of, you know,

30:12SPEAKERit's aware that if there's a thing that's pink here

30:16SPEAKERand there's a thing that's black here

30:17SPEAKERon the other side of the screen,

30:18SPEAKERoften the things on this will be this and this and this.

30:21SPEAKERAnd the upshot is you get fewer weird artifacts.

30:24SPEAKERThere's still gonna be artifacts,

30:25SPEAKERbut they're novel and new in the same way

30:29SPEAKERthat like when you change from one kind of video compression

30:33SPEAKERto another kind of video compression,

30:35SPEAKERthere's still artifacts, but they're different

30:37SPEAKERand your brain doesn't pick them up as easily.

30:39SPEAKER- Right, so is transformer a generic

30:41SPEAKERkind of machine learning term of art then?

30:43SPEAKER'Cause Jensen dropped that in the reveal at CES for this.

30:47SPEAKERAnd I mean, it's not really the place for it on stage,

30:50SPEAKERbut he didn't really do a great job

30:51SPEAKERof defining what a transformer is or means.

30:54SPEAKERBut the point is that's not a,

30:55SPEAKERthis is not some like new NVIDIA innovation.

30:58SPEAKERThis is something that exists in machine learning generally.

31:00SPEAKER- Yeah, this is something that came out of research

31:03SPEAKERlike 10, almost 10 years ago at this point.

31:05SPEAKER- Okay.

31:07SPEAKER- And yeah, the idea is more context.

31:10SPEAKERLike I said, the important thing to know in this case

31:13SPEAKERis for DLSS, the convolutional networks,

31:16SPEAKERneural nets are aware of the pixels

31:18SPEAKERin the immediate area around each pixel

31:20SPEAKERthat it's generating with the upsampling.

31:22SPEAKERWith the transformer, it's more aware of the broad pictures

31:26SPEAKERacross the entire image.

31:27SPEAKER- Right, and then just to reiterate

31:29SPEAKERin case it's gotten lost here,

31:30SPEAKERthe convolutional neural net, the CNN

31:33SPEAKERis the old way that DLSS did upscaling

31:35SPEAKERand the transformer is essentially just the new way.

31:38SPEAKERI'm seeing here there's a performance difference.

31:41SPEAKER- There is a little bit of a performance difference.

31:42SPEAKERThe transformer is more expensive.

31:44SPEAKERI've done benchmarks.

31:45SPEAKERI have benchmarks running on fully path traced cyberpunk

31:50SPEAKERrunning on the entire stack that runs DLSS for real

31:54SPEAKERfrom the 22080, of the high end stack rather,

32:00SPEAKERthe 20 series, 30 series, 40 series, and 50 series.

32:03SPEAKERAnd it seems to be between a two and 3%,

32:06SPEAKERmaybe four at the absolute outside performance hit.

32:09SPEAKER- That's not too bad. - To run the transformer

32:10SPEAKERmodel, it's in beta now.

32:12SPEAKERSo it's only available on a couple of games

32:19SPEAKERfor us as pre-release.

32:20SPEAKER- Okay, I was gonna ask how you kind of opted in

32:22SPEAKERif games are gonna, and surely games are not going

32:25SPEAKERto expose like different types of DLSS upscaling

32:28SPEAKERat some point, they'll just switch out.

32:30SPEAKERI mean, that just seems like needless complexity

32:32SPEAKERfor the average person to have to figure out

32:34SPEAKERwhat to set it to.

32:36SPEAKER- Yeah, so for us, there's a flag in cyberpunk

32:39SPEAKERthat you can switch between CNN and transformer

32:41SPEAKERon the DLSS panel.

32:42SPEAKERI don't know if that's gonna ship.

32:44SPEAKERI looked, like you said, I hope not

32:46SPEAKER'cause it's kind of confusing for normal people.

32:49SPEAKERThere is a flag in the Nvidia app

32:51SPEAKERin what used to be the kind of graphics settings pane

32:54SPEAKERfor in the old Nvidia control panel

32:56SPEAKERwhere you can flip it on for the games that support it.

32:59SPEAKERAnd right now it's available in cyberpunk in Vale Guard,

33:03SPEAKERDragon Age Vale Guard.

33:05SPEAKERI did my testing in cyberpunk just 'cause it has

33:07SPEAKERa benchmark loop and it's easy to get

33:10SPEAKERfairly repeatable results.

33:12SPEAKER- Just to make sure I understand,

33:13SPEAKERI mean, I think you said this,

33:14SPEAKERbut so this new version of DLSS upscaling

33:16SPEAKERis available on basically every card that does DLSS, right?

33:20SPEAKERAll the way back to the 20 series?

33:23SPEAKER- Yeah, it should, but I mean,

33:25SPEAKERwith the beta drivers that they gave us for testing, yes.

33:29SPEAKERI don't know what the rollout's gonna be.

33:30SPEAKERI assume that this is gonna be available

33:32SPEAKERfor everybody to try on the 30th with,

33:35SPEAKERbut what the rollout, so they block the transformer model

33:39SPEAKERon games that don't actively support in the driver panel.

33:42SPEAKER- That makes sense, but the big point here though

33:44SPEAKERis this is definitely not a 50 series exclusive

33:47SPEAKERor even limited to like 40 and newer.

33:50SPEAKER- Yeah, there's actually hardly any of the stuff

33:52SPEAKERthat we're gonna talk about,

33:53SPEAKERexcept for the multi-frame gen is limited to 50 series.

33:56SPEAKERAlmost everything is coming to the older cards.

34:00SPEAKERAnd yeah, the places that I noticed it working better

34:05SPEAKERare on things like chain link fences and transparency

34:09SPEAKERthat the convolutional networks

34:12SPEAKERhave traditionally had problems with.

34:16SPEAKERI'm gonna say having captured a bunch of footage

34:20SPEAKERand done frame by frame comparisons

34:22SPEAKERof eight different ways of rendering this stuff,

34:28SPEAKERthe effect's real subtle.

34:32SPEAKERThere's an occasional glitch that you'll notice

34:34SPEAKERthat jumps out at you, but for the most part,

34:37SPEAKERyou have to be resting your eyeballs on the screen

34:39SPEAKERand moving the mouse real fast to see most of this stuff.

34:41SPEAKER- Like you said earlier, at 120 Hertz or more,

34:44SPEAKERit's kind of hard to pick out weird artifacts

34:46SPEAKERand individual frames or even a bundle of frames.

34:48SPEAKERSo like, I'm not expecting, you know,

34:50SPEAKERI mean, once this option rolls out everywhere,

34:52SPEAKERlike of course for the tiny performance hit,

34:54SPEAKERI'll probably flip on the better version,

34:56SPEAKERbut like, is it actually gonna be noticeable

34:58SPEAKERin a profound way?

34:59SPEAKERProbably not.

35:01SPEAKER- I think actually, like to me,

35:04SPEAKERhaving spent, I spent two days playing games

35:07SPEAKERwith the stuff on and off and trying different ways,

35:10SPEAKERboth qualitatively analyzing and just experiencing.

35:14SPEAKERAnd for me, that metaphor, the analog is video compression,

35:18SPEAKERwhere when you move from MPEG-2 to H.264,

35:26SPEAKERthere's still compression artifacts in H.264,

35:29SPEAKERbut they're different looking than the MPEG-2 ones.

35:33SPEAKERAnd it was a little jarring.

35:35SPEAKERNow, of course, when you go back and watch an MPEG-2 video

35:37SPEAKERfrom like the nineties, it looks real blocky and weird

35:40SPEAKERbecause there's block artifacts in those videos

35:44SPEAKERthat we just filter, our brains filtered out

35:46SPEAKER'cause we were so used to seeing them.

35:47SPEAKERAnd I think this is gonna be the same situation.

35:49SPEAKER- Yeah, I mean, this is a worthy goal to be clear,

35:50SPEAKEReven if the differences are not hugely noticeable,

35:52SPEAKERlike in any situation where you're conjuring data

35:56SPEAKERfrom nothing, you know, and you're extrapolating,

35:59SPEAKERyou're extrapolating frames that don't exist

36:01SPEAKERfrom smaller ones that do, you know,

36:05SPEAKERaccuracy is a worthy pursuit, right?

36:08SPEAKERLike they should always be striving

36:10SPEAKERto make this more accurate,

36:11SPEAKEReven if the differences are very subtle.

36:14SPEAKER- Well, and the benefit of doing this

36:16SPEAKERis that you have less stuff going through

36:19SPEAKERthe raster pipeline, which, let's see,

36:23SPEAKERthe way to put this is performance hiccups

36:25SPEAKERin the raster pipeline,

36:26SPEAKERwhich is where all of this stuff starts,

36:30SPEAKERhave a negative downstream consequences

36:33SPEAKERon everything else in the pipeline.

36:35SPEAKERSo, you know, your brain is really,

36:38SPEAKERsome people, we should start,

36:41SPEAKERthere's a cognitive spectrum for people

36:47SPEAKERwith regards to things like frame rate and micro stutter,

36:50SPEAKERand, you know, frame rate is the number of frames

36:53SPEAKERthat are thrown on the screen every second.

36:54SPEAKERMicro stutter happens when one or two frames in a row

36:59SPEAKERtake a lot more time, you know, double or triple the time

37:04SPEAKERto render than the frames on either side of them.

37:06SPEAKERSo it feels like you have a hitch.

37:08SPEAKERAnd in an extreme case, like I had a driver bug

37:11SPEAKERthe other day when I was testing something,

37:13SPEAKERthat I was getting 1,000 frames per second

37:16SPEAKERaverage frame rate, but I was also having some,

37:20SPEAKERmy 1% low frames, the slowest frames,

37:22SPEAKERI were running at like eight,

37:24SPEAKERthe equivalent of eight frames a second.

37:26SPEAKERAnd even though the vast majority, 94% of those frames

37:30SPEAKERwere running at 1,000 frames per second,

37:33SPEAKERit felt like the game was running at 10 frames a second

37:35SPEAKERbecause the stutter on those hold frames,

37:38SPEAKERyour brain perceived those at such a high level

37:40SPEAKERthat the game felt really, really slow.

37:42SPEAKER- Right.

37:44SPEAKER- So a lot of the work that they do here

37:47SPEAKERis to reduce that micro stutter

37:49SPEAKERand to smooth out the frame rate

37:50SPEAKERand get the frame pacing really good.

37:52SPEAKERIt's why, my favorite example is when you play Doom,

37:56SPEAKERone of the new modern Doom games, 2016 or Eternal.

37:59SPEAKERAnd those games are fast, but they're not like,

38:01SPEAKERthey run at like 90 frames a second,

38:03SPEAKERbut they feel so much smoother and faster than most games.

38:07SPEAKERAnd it's because the team that works on those

38:09SPEAKERand the engine they use, they pay really attention.

38:12SPEAKERThey pay a lot of attention to frame pacing.

38:14SPEAKERSo the frames come at a really even rate.

38:16SPEAKERSo you're 90 frames a second,

38:18SPEAKERyour minimum lows are also like 88 frames a second

38:20SPEAKERor something like that, typically on that game.

38:23SPEAKER- Yeah, like id can't, modern id cannot get enough credit

38:25SPEAKERfor how much they prioritize smoothness

38:28SPEAKERand presentation and efficiency in their tech.

38:30SPEAKER- Yeah, and it makes their games feel like they're,

38:33SPEAKERit makes their games feel like they're coming

38:34SPEAKERfrom a different era, they're so fast.

38:36SPEAKER(laughing)

38:37SPEAKER- I mean, it feels like butter to play those games

38:39SPEAKERis the best way to put it.

38:41SPEAKER- So, okay, so which brings us to multi-frame gen.

38:44SPEAKERFrame gen is one of those things

38:45SPEAKERthat Nvidia rolled out last generation.

38:49SPEAKERAnd they specifically said, "Hey, this is a way

38:51SPEAKERthat we're gonna help combat micro stutter

38:53SPEAKERto give you some machine generated frames

38:56SPEAKERto fill in gaps when there's timing problems

38:59SPEAKERwith your raster pipeline or your ray tracing pipeline,

39:03SPEAKEReither one."

39:05SPEAKERThose timing gaps, like if you've played Star Wars Outlaws,

39:08SPEAKERI don't, apologies for picking on Massive,

39:10SPEAKERbut that game has some crazy micro stutter problems

39:13SPEAKERin my experience.

39:14SPEAKERAnd like when you flip on frame gen in that game,

39:17SPEAKERit feels much, much better

39:19SPEAKERbecause their frame pacing is bad.

39:22SPEAKERAnd so last gen they rolled out one raster frame,

39:29SPEAKERone ray traced frame, and then one,

39:32SPEAKERlike you take one generated frame traditionally,

39:34SPEAKERand then it would machine generate another one

39:36SPEAKERusing the tensor cores on the GPU.

39:37SPEAKER- Like forever, I hesitate to use these terms

39:40SPEAKERbecause there's a lot of what, you know,

39:42SPEAKERwhat I'm about to say.

39:43SPEAKER- There's a lot of baggage.

39:44SPEAKER- There's a lot of the FUD, I don't know what the term is.

39:47SPEAKERThere's a lot of Sturm und Drang out there

39:50SPEAKERabout this feature.

39:51SPEAKER- There's a lot of feelings about AI generated frames.

39:54SPEAKER- The term I hesitate to use, but it is descriptive

39:56SPEAKERis I was going to say for every real frame,

39:58SPEAKERthere's an AI frame.

39:59SPEAKERIn the original single frame gen situation,

40:03SPEAKERyou get one real rendered frame

40:05SPEAKERand then one machine generated frame.

40:07SPEAKER- Yeah.

40:07SPEAKERYeah, and so the big thing with this generation of cards

40:11SPEAKERis that they've done something

40:12SPEAKERthat sounded crazy in the keynote.

40:14SPEAKERI think we both reacted about the same way,

40:15SPEAKERwhich is they went from one real frame

40:19SPEAKERgenerating one AI generated frame to having one real frame

40:22SPEAKERgenerating three additional AI frames.

40:25SPEAKER- Yeah.

40:28SPEAKER- Having, okay, so I'm gonna not bury the lead.

40:30SPEAKERHaving spent a lot of time playing with this,

40:32SPEAKERit feels really fricking good.

40:34SPEAKER- Really?

40:35SPEAKER- When you flip it on.

40:36SPEAKER- Really?

40:37SPEAKER- Yeah, I was, so I was a little cool

40:40SPEAKERon the first gen baked frames.

40:42SPEAKER- Yes.

40:43SPEAKER- AI generated frames.

40:44SPEAKER- I'm not a fan.

40:45SPEAKER- Just 'cause the latency increase is pretty substantial

40:48SPEAKERon most of them.

40:50SPEAKERThe meeting, the time from when you move your mouse

40:54SPEAKERto when you feel that movement happen on screen felt mucky.

40:58SPEAKERWith a controller, you notice it a little bit less

41:00SPEAKER'cause the controller doesn't do one-to-one input

41:01SPEAKERfor the most part.

41:02SPEAKERAnd so it kind of, your mushiness gets lost

41:07SPEAKERin the mushiness of the controller.

41:10SPEAKERWith the mouse and keyboard,

41:11SPEAKERespecially at a mid to high sensitivity,

41:13SPEAKERI noticed it immediately and it felt like you were kind of,

41:15SPEAKERit felt like you had weight on your mouse

41:18SPEAKERand put in a way that feels bad.

41:20SPEAKER- That's a good way to describe it, yes.

41:22SPEAKER- This, the multi-frame gen on Outlaws

41:24SPEAKERmakes Outlaws feel like Doom Eternal.

41:26SPEAKER- Wow, that's a pretty bold statement.

41:30SPEAKER- Yeah.

41:31SPEAKER- All right.

41:31SPEAKER- I mean, it feels, playing with a mouse and keyboard,

41:34SPEAKERjust to be clear, I played Outlaws with frame gen on

41:37SPEAKERand a game pad because the mouse input felt so swimmy

41:40SPEAKERand the micro stutter was annoying

41:43SPEAKERto the point that I didn't play it without that turned on.

41:45SPEAKER- Well, I mean, that's effectively a console game.

41:46SPEAKERLike I could see, controller is probably the ideal way

41:49SPEAKERto play that game regardless of other factors.

41:52SPEAKER- Yeah.

41:54SPEAKER- I'm not in any way doubting what you're saying,

41:55SPEAKERbut I am going to say,

41:56SPEAKERI'm going to continue to be very skeptical of this feature

41:58SPEAKERuntil I try it for myself and come away feeling

42:01SPEAKERthe same way you do because it sounds too good to be true,

42:03SPEAKERbut go on.

42:04SPEAKER- A hundred percent.

42:05SPEAKERSo the thing to note, the thing to know is that,

42:08SPEAKERand I think Digital Foundry said this first years ago

42:11SPEAKERwith the first gen is that this isn't tech

42:14SPEAKERthat gets you from 20 frames a second to 40 frames a second

42:18SPEAKERor 80 frames a second.

42:20SPEAKERThis is tech that gets you from 60 frames per second

42:23SPEAKERto the refresh rate of your 240 Hertz monitor, right?

42:27SPEAKER- So you're saying you need a large sample size,

42:29SPEAKERa large pool of real frames for this to work right?

42:32SPEAKER- Just as DLSS, which gets better when you start

42:35SPEAKERwith a 1080p or a 960 or 1200 horizontal resolution

42:41SPEAKERor vertical resolution rather image,

42:43SPEAKERand then scale that up so you have more data to start with.

42:46SPEAKERThis works better when you have more frames to work with.

42:49SPEAKERWhen I tried it with a game that was running at 20,

42:52SPEAKERwhen I cranked everything up

42:53SPEAKERand ran it at a really high resolution

42:55SPEAKERto get the game to run at 20 frames per second natively,

42:58SPEAKERand then turned on multi-frame gen,

42:59SPEAKERit felt swimmy and crappy just like it did before.

43:02SPEAKER- Really, even with the multi-frame,

43:03SPEAKER'cause that was, I mean, when they first rolled this out,

43:06SPEAKERI'm trying to think frame gen was on the 30 cards, right?

43:09SPEAKER- 40 series was when frame gen was introduced, yeah.

43:12SPEAKER- Anyway, that was the original problem though,

43:13SPEAKERwas you were on hardware that,

43:15SPEAKERlike with the paths racing in Cyberpunk, for example,

43:17SPEAKERbarely could manage 20, 30, maybe 40 frames a second.

43:21SPEAKERSo you were getting the swiminess

43:23SPEAKERbecause you had so few,

43:25SPEAKERyou had so little actual performance to work with.

43:27SPEAKER- So there's a couple of things that they changed

43:29SPEAKERthis go round.

43:30SPEAKERAs before, it requires reflex to work.

43:33SPEAKERSo the game has to support DLSS multi-frame gen and reflex.

43:39SPEAKERReflex provides vectors, reflex is a low latency technology.

43:44SPEAKERWe'll talk more about it in a minute.

43:45SPEAKERIt's designed on its face to reduce the time

43:48SPEAKERfrom motion of a mouse or an input device

43:52SPEAKERto the photons on the screen,

43:55SPEAKERreflecting that movement that the mouse does.

43:58SPEAKERBut as a side effect of how it works,

44:00SPEAKERit has to provide vectors for how the camera

44:02SPEAKERand objects in the game are moving to the video card driver.

44:07SPEAKERAnd multi-frame gen use that vector data

44:10SPEAKERto generate the next round of frames.

44:12SPEAKERSo it's not just looking at this picture

44:15SPEAKERand saying, oh, this picture,

44:17SPEAKERwe're gonna make this picture look like the next picture.

44:18SPEAKERIt's saying, oh, okay, the input on the user

44:22SPEAKERand the input on the onscreen objects

44:24SPEAKERare moving in these directions.

44:26SPEAKERHere's what we can infer from that,

44:28SPEAKERrendering the next frame from that.

44:29SPEAKER- Yeah, it's a very educated guess.

44:31SPEAKERIt's not just working off of an image and nothing else.

44:35SPEAKER- Yeah, so the other thing is that

44:36SPEAKERthe multi-frame gen, the five series cards

44:40SPEAKERmove to the frame spacing, frame pacing

44:42SPEAKERfrom the CPU to the hardware on the GPU now.

44:46SPEAKERSo they have more control over when that happens,

44:49SPEAKERit's less likely to get bottlenecked,

44:50SPEAKERand it reduces the time from when the frame is generated

44:54SPEAKERto when it's displayed on the screen.

44:55SPEAKER- Okay.

44:56SPEAKER- They're also running, all three of the frames

45:01SPEAKERare generated off of the one raster frame at the same time.

45:05SPEAKER- Oh.

45:06SPEAKER- And then, so they're not generating a frame

45:08SPEAKERand then generating the next frame off of the AI,

45:11SPEAKERfirst AI frame and the third frame

45:13SPEAKERoff of the second AI frame.

45:14SPEAKER- Oh, interesting.

45:15SPEAKER- So they generate one and then they warp the frame

45:18SPEAKERto match the input that you've done at the last second

45:22SPEAKERright before it's rendered, basically.

45:24SPEAKER- Also through machine learning?

45:25SPEAKER- Also through machine learning, yeah.

45:27SPEAKER- So abstract and hard to get your head around.

45:31SPEAKER- The way I think of it is it's kind of like anti-aliasing,

45:34SPEAKERwhich anti-aliasing would render some or all of the screen

45:38SPEAKERat a higher resolution and then shrink it down

45:40SPEAKERto reduce the jaggy aliasing on like diagonal stuff

45:45SPEAKERback in the day.

45:46SPEAKER- Is that, that's super sampling, right?

45:48SPEAKER- Well, that's super sampling is a type of anti-aliasing.

45:51SPEAKERSo you render up and then you shrink it back down

45:53SPEAKERand it blurs the pixels together.

45:55SPEAKERSo there are various types of anti-aliasing

45:59SPEAKERthat will just focus on the parts of the screen

46:01SPEAKERthat are tended, that it picks up jaggies on.

46:07SPEAKERThere's temporal AA, which blurs frame to frame

46:11SPEAKERso that it works forward in time, basically.

46:15SPEAKERBut this is kind of like AA for time

46:19SPEAKERin that it is just giving you additional frames

46:22SPEAKERfor when the rendered frames come in hot or slow.

46:26SPEAKERAnd yeah, there's only a handful of games

46:30SPEAKERthat it works with right now.

46:30SPEAKERCyberpunk is one of them, Outlaws is one of them.

46:33SPEAKERI think Alan Wake does.

46:34SPEAKERIt was one of the other games that they offered us to test.

46:38SPEAKERBut I can't, actually, I didn't,

46:40SPEAKERthat was the thing I spent the least time with.

46:43SPEAKERThe Outlaws change was pretty remarkable.

46:47SPEAKERCyberpunk, Cyberpunk runs great.

46:49SPEAKERLike just to be clear, I actually,

46:52SPEAKERI don't know what the number is with all the path tracing on

46:54SPEAKERand no DLSS or no frame gen, but like Cyberpunk

46:59SPEAKERon a 50/90 at 4K will run locked 30 dead solid

47:04SPEAKERwith everything turned all the way up and no AI stuff.

47:07SPEAKER- Native 4K?

47:09SPEAKER- Native 4K. - With no assists?

47:10SPEAKEROkay. - Yeah, no assists.

47:13SPEAKERThe multi-frame gen, I expect it to still be controversial.

47:17SPEAKERSome people in the first round of reviews didn't even test it

47:20SPEAKERwhich I get like, you know, whatever, do your thing.

47:23SPEAKERI feel like it's pretty, like we've gone from a point,

47:28SPEAKERto me, this feels like the DLSS2 moment.

47:31SPEAKERYou know, DLSS1 was trained per game to do the upsampling.

47:36SPEAKERSo you had to, you know, they had to train each,

47:38SPEAKERthey had to train a different model

47:39SPEAKERfor each game, basically.

47:41SPEAKERDLSS2 was when they rolled out the generalized model.

47:44SPEAKERAnd that felt like a moment when,

47:46SPEAKERwhen DLSS suddenly became a viable tool.

47:49SPEAKERThis feels like the moment when the frame gen stuff

47:51SPEAKERbecomes worth turning on.

47:52SPEAKER- Okay. - To me.

47:53SPEAKER- All right, I will approach with cautious optimism

47:55SPEAKERbut I definitely still need to be convinced firsthand.

47:58SPEAKER- I think the hard part of this is

48:01SPEAKERyou have to play it to understand it.

48:03SPEAKERLike you, like we can't, if we were on YouTube,

48:06SPEAKERwe couldn't show you a video

48:07SPEAKERthat's gonna make you understand why this is good.

48:09SPEAKERIt's purely a difference in feel.

48:12SPEAKERAnd like I said, you're not gonna go

48:15SPEAKERfrom 20 frames a second to 80 frames a second.

48:17SPEAKERYou're gonna go from 60 or 80 frames a second

48:19SPEAKERto 240 frames a second.

48:21SPEAKER- Or 144.

48:23SPEAKER- Yeah, my big question is

48:26SPEAKERhow is this gonna work on the cards

48:28SPEAKERthat are less performant at the lower end?

48:31SPEAKER'Cause you know, Jensen came out on the stage

48:34SPEAKERand was like, hey, your 4070 is gonna be as fast,

48:36SPEAKERyour 5070 is gonna be as fast as a 4090,

48:39SPEAKERwhich is true from a certain point of view

48:43SPEAKERif you have on the 75 titles that will support DLSS4

48:46SPEAKERright out of the gate.

48:47SPEAKER- Yeah, and world's biggest asterisk there

48:49SPEAKERbecause who knows how that'll feel

48:50SPEAKERif those cards are turning out fewer native frames

48:54SPEAKERthan you may get the swimminess

48:55SPEAKERthat makes that not super worth using.

48:58SPEAKER- That's it, that's exactly it.

48:59SPEAKER- Real quick before we move on,

49:00SPEAKERjust a quick addendum to something I mentioned earlier.

49:03SPEAKERApparently Digital Foundry did an interview at CES

49:07SPEAKERwith Brian Catanzaro.

49:09SPEAKER- Yeah, it's a fantastic video if you haven't watched it.

49:12SPEAKER- VP of Applied Deep Learning Research.

49:14SPEAKERAnyway, he hinted that frame generation

49:16SPEAKERmight come to the 30 cards at some point.

49:18SPEAKER- All the way back to the 30 cards?

49:20SPEAKER- They're not committing to it, but.

49:22SPEAKER- Oh, well, actually, I mean, the actual quote from Catanzaro

49:25SPEAKERis maybe a little more general here.

49:27SPEAKERI asked him about getting frame generation

49:29SPEAKERworking on previous cards.

49:30SPEAKERI think this is primarily a question of optimization

49:32SPEAKERand also engineering, and then the ultimate user experience.

49:35SPEAKERSo we're launching this frame generation,

49:37SPEAKERthe best multi-frame generation technology with 50 series,

49:40SPEAKERand we'll see what we're able to squeeze

49:42SPEAKERout of older hardware in the future, so.

49:44SPEAKER- That to me says maybe multi-frame

49:46SPEAKERwill come to the 40 series.

49:47SPEAKER- That is exactly my conclusion

49:49SPEAKERfrom reading the actual quote here.

49:50SPEAKERMaybe frame generation on 30 is a bit more of a stretch

49:55SPEAKERbased on that quote, but they're ruling out

49:58SPEAKERsome improvements on old hardware in the future.

50:01SPEAKER- Look, I am a fan of this work.

50:07SPEAKERI am excited to see if it comes to the older hardware.

50:10SPEAKERI was pleasantly surprised by the transformer model

50:14SPEAKERrunning on the older Tensor Cores.

50:15SPEAKERLike, it ran fine on the 2080.

50:18SPEAKERNow, was it getting a playable frame rate

50:20SPEAKERin path traced Cyberpunk on a 2080?

50:23SPEAKERNo. - Yeah.

50:24SPEAKER- But the performance drop for turning on transformer model

50:27SPEAKERwas the same pretty much across the board.

50:30SPEAKERWell, let's see, what else we got?

50:31SPEAKERReflex 2, Reflex 2 is not available to test right now,

50:34SPEAKERso I don't know anything about Reflex 2.

50:36SPEAKER- What? - Yeah.

50:37SPEAKERIt's only in the finals in Valorant.

50:38SPEAKER- I'm sorry, I thought the whole pitch

50:41SPEAKERwas that Reflex 2 was a big part

50:43SPEAKERof how they were making multi-frame gen palatable.

50:46SPEAKER- That is fascinating that you say that,

50:47SPEAKER'cause I thought the exact same thing

50:49SPEAKERfrom the keynote.

50:50SPEAKERBut it is, they might be using frame warp

50:54SPEAKERfor the MFG stuff for those last frames.

50:57SPEAKERThat was the part that I think was confusing to people,

51:00SPEAKERwhich is also being used in Reflex 2.

51:03SPEAKERAgain, Reflex is the, hey,

51:05SPEAKERreduce the time it takes to move your mouse.

51:07SPEAKERTypically, especially for players that play

51:10SPEAKERat medium or high sensitivity and high frame rate,

51:13SPEAKERlike the render position of the game

51:17SPEAKERis typically one frame or more behind

51:20SPEAKERwhere the mouse actually is for the person controlling it.

51:24SPEAKERGenerally speaking, people who are really good

51:26SPEAKERat first-person shooters, like professional players

51:28SPEAKERwho train, drill to learn how to bypass this effect.

51:32SPEAKER- Oh, wow.

51:34SPEAKER- So they'll move the mouse and then stop a frame early.

51:38SPEAKERBut the whole point of Reflex is that you don't have

51:41SPEAKERto necessarily feel those hitches.

51:46SPEAKERSo Reflex 2 is adding frame warp,

51:49SPEAKERwhich is basically they take the frame

51:53SPEAKERthat they get from the rendering pipeline.

51:55SPEAKERThis is all with frame generation, all that stuff off,

51:57SPEAKER'cause you don't want frame generation

51:59SPEAKERon competitive games at all.

52:02SPEAKERSo the way the Reflex 2 works is they've got

52:07SPEAKERa low-latency frame generator or infill generator, basically.

52:12SPEAKERSo they warp the camera position based on the vectors

52:15SPEAKERthey get from Reflex at the last possible second,

52:18SPEAKERand then in-paint in any stuff

52:19SPEAKERthat doesn't render with the AI.

52:23SPEAKER- Seems so ridiculous.

52:24SPEAKERI think I've said this before.

52:26SPEAKERWhen I just visualize this process mentally,

52:28SPEAKERit seems so seat of the pants, absolute last second.

52:33SPEAKERThe fact that you're literally distorting the frame

52:35SPEAKERand then having to paint around the edges

52:37SPEAKERthat you no longer have just to get the frame out,

52:39SPEAKERit's felt like it just seems so crazy in concept,

52:41SPEAKERbut I guess clearly they have made it work.

52:43SPEAKER- Yeah, so I mean, like I said, I haven't seen it running.

52:48SPEAKERThe folks who did see it running,

52:49SPEAKERI think DF did a video about it at CES

52:51SPEAKER'cause they had the finals demo running there,

52:53SPEAKERa digital foundry.

52:57SPEAKERI'm wildly curious to play this.

53:00SPEAKERIt's only coming to the finals in Valorant at first,

53:03SPEAKERavailable for 50 series GPUs.

53:06SPEAKERThey're saying 75 to 50% faster lower latency

53:11SPEAKERthan reflex one with boost to the high speed reflex one.

53:15SPEAKERSo I'm interested to see that.

53:17SPEAKERThat means that your latency will be down to,

53:19SPEAKERI think they said two milliseconds in Valorant.

53:21SPEAKER- Wow.

53:22SPEAKER- Which is like, at that point, they're good.

53:24SPEAKERThey don't have to worry about this anymore.

53:25SPEAKER- That's a solved problem at that point.

53:27SPEAKERIs reflex two coming to every card?

53:30SPEAKER- Starting with the 50 series,

53:32SPEAKERbut it seems like it's coming to the older cards later.

53:34SPEAKER- Okay. - Yeah.

53:35SPEAKER- Okay.

53:36SPEAKER- My guess is that the low latency fill in,

53:39SPEAKERthey have to do optimization

53:41SPEAKERto get working on the older cards.

53:42SPEAKER- Yeah, yeah, I mean, I guess the good news,

53:44SPEAKERI mean, even if the entirety of reflex two

53:47SPEAKERis not required to make multi-frame gen work,

53:49SPEAKEReventually once both of those things

53:51SPEAKERare fully supported in games,

53:53SPEAKERthey will work in concert together

53:54SPEAKERto hopefully make multi-frame gen better.

53:58SPEAKER- So I actually, the multi-frame gen,

54:00SPEAKERI don't know if you're ever gonna wanna turn on

54:02SPEAKERcompetitive games.

54:03SPEAKER- Oh yeah, definitely, absolutely not.

54:04SPEAKER- 'Cause you don't wanna add latency

54:05SPEAKERand you don't wanna add it,

54:06SPEAKERlike you want real frames ground truth

54:08SPEAKERas close as you can get, right?

54:10SPEAKER- Yeah, but most people are not making their living

54:12SPEAKERplaying first person shooters, so that's, we'll see.

54:16SPEAKER- So I did a little bit of informal testing

54:18SPEAKERwhere I had my daughter come in and flip multi-frame,

54:20SPEAKERor frame gen on in COD for me a week or so ago,

54:24SPEAKERand she would flip it on and off

54:26SPEAKERwithout me knowing which one it was,

54:27SPEAKERand then I would play a handful of games

54:29SPEAKERon a game that I have a pretty level KD ratio on,

54:34SPEAKERon a game type and level that I have a straight 2.0 usually,

54:39SPEAKERand I would drop to 1.5 to 1.6 when frame gen was on,

54:45SPEAKERpretty consistently.

54:46SPEAKER- The old Coke and Pepsi blind taste test.

54:49SPEAKER- Yeah, except for the Pepsi thing only worked

54:53SPEAKERbecause they only let you take one sip

54:55SPEAKERand everybody likes Pepsi when it's one sip, it turns out,

54:57SPEAKER'cause it's real sweet.

54:58SPEAKER- Arsh, man.

55:00SPEAKER- Look, Pepsi sucks, that's the official policy

55:02SPEAKERof this podcast, Brad, I'm sorry.

55:04SPEAKER- Arsh, arsh.

55:05SPEAKER- You get your lightning plug, I get Pepsi sucks.

55:07SPEAKER- All right.

55:08SPEAKER- Worth mentioning that the Reflex2 stuff,

55:10SPEAKERit works in CPU bottleneck scenarios

55:13SPEAKERwhere Reflex1 is not as effective.

55:14SPEAKER- Oh, okay.

55:17SPEAKER- And then I think we've talked about this already,

55:19SPEAKERthe stuff that works on the old cards

55:21SPEAKERversus what works on new cards,

55:22SPEAKERit's pretty much the same as before.

55:24SPEAKERFrame gen is four series, multi-frame gen is five series,

55:27SPEAKEReverything else works on everything except for maybe Reflex2,

55:31SPEAKERwhich seems like it'll start on five series

55:33SPEAKERand then roll out to the older cards over time.

55:35SPEAKERSo that's like- - Good news for, yeah,

55:39SPEAKERgood news for people who have old cards.

55:40SPEAKERNice to see new features coming to this stuff

55:43SPEAKERand particularly like 40 series

55:44SPEAKERwhere the performance increase on these is noticeable

55:48SPEAKERbut not profound, to get just about all of the new features

55:52SPEAKERon a 40 card is pretty good news.

55:54SPEAKER- Yeah, I mean, I think the classic,

55:57SPEAKERhey, buy every other generation,

55:58SPEAKERif you're gonna upgrade often, it remains safe, right?

56:01SPEAKER- Yeah.

56:02SPEAKER- And 40 series will hold through this two years

56:05SPEAKERif that's what you're worried about.

56:08SPEAKERBut I think this is, if you are a,

56:11SPEAKERI always buy the new card,

56:13SPEAKERthere's stuff here that's worth buying.

56:14SPEAKERThe question to me is whether to buy an AIB third-party card

56:18SPEAKERwith a more traditional cooler

56:20SPEAKERor go with the smaller, potentially hotter two-slot.

56:23SPEAKERAnd unfortunately, the reviews for the Attenborough coolers,

56:26SPEAKERthat embargo dropped as we were recording this podcast

56:29SPEAKERand they're all over the place depending on,

56:33SPEAKERalso the price on the add-in board cards

56:35SPEAKERis all over the place as well,

56:37SPEAKERlike up to $2,800 in some cases it looks like.

56:40SPEAKERSo yeah, there's a couple of other things

56:43SPEAKERI wanna hit real quick if we can.

56:45SPEAKERThey demoed at the keynote some new features

56:51SPEAKERand this is the thing they often do.

56:53SPEAKERIt's usually a mixture of stuff

56:55SPEAKERthat's actually the future of computer rendering for games

56:58SPEAKERand then some hui. (laughing)

57:00SPEAKER- Seems like a good way to put it

57:01SPEAKER'cause some of this stuff seems like

57:03SPEAKERit will become bedrock practice in the next five to 10 years

57:08SPEAKERand some of it seems like, no, thank you.

57:10SPEAKER- Yeah, the neural shaders is definitely a,

57:12SPEAKERhey, this is gonna be bedrock in the future.

57:14SPEAKER- Absolutely, absolutely of this list.

57:16SPEAKERThat seems like the thing that like, okay,

57:17SPEAKERnext year people are gonna be using this

57:19SPEAKERto do interesting things.

57:20SPEAKER- Yeah, so this uses AI modeling to basically,

57:25SPEAKERit kind of, it both provides a performance

57:31SPEAKERand a functionality increase by using AI,

57:36SPEAKERwhat's the right way to say this?

57:37SPEAKERBy generating materials in games using traditional methods

57:41SPEAKERand then kind of chunking that down into an AI shader.

57:45SPEAKERSo if a material in a game,

57:46SPEAKERlike any surface you look at in a game

57:48SPEAKERhas a material attached to it

57:48SPEAKERand they can have a whole bunch of stuff,

57:50SPEAKERwhole bunch of ideas attached to it.

57:52SPEAKERIt can have the color of the texture,

57:54SPEAKERit can have the texture of the surface,

57:57SPEAKERlike whether it's bumpy or smooth or shiny or-

58:00SPEAKER- Smoothness, opacity, translucency.

58:02SPEAKER- Emissiveness.

58:03SPEAKER- Yeah, and even down to like things like fur

58:06SPEAKERand kind of simulating more complex things on the surface.

58:12SPEAKER- Whether there's subsurface scattering,

58:16SPEAKERwhether there's materials that appear

58:19SPEAKERthrough the translucency, like skin.

58:21SPEAKERIf you look at your hand and you're in a bright light,

58:25SPEAKERyou can see the veins and the dark spots

58:29SPEAKERwhere the bones are underneath, you can see the pores,

58:31SPEAKERyou can see all this different stuff.

58:33SPEAKERAnd a lot of that happens because the light goes

58:34SPEAKERinto the top layers of your skin,

58:35SPEAKERbounces around a little bit and then comes back out.

58:38SPEAKERAnd so the benefit of neural shaders is, yeah,

58:40SPEAKERyou can do that stuff with traditional shader model stuff,

58:43SPEAKERbut it's really computationally and memory expensive.

58:47SPEAKERIt costs a lot of memory, it costs a lot of computation.

58:50SPEAKERAnd what they're doing is they're applying machine learning

58:53SPEAKERto that space in a constrained way.

58:56SPEAKERSo they're taking the models

58:58SPEAKERthat are generated traditionally.

59:00SPEAKERAnd it's almost like, it's not exactly compression

59:03SPEAKERbecause it's doing more than that,

59:04SPEAKERbut it's like compressing the computation down

59:07SPEAKERby using machine learning, which is like,

59:09SPEAKERI think everybody who listens to this podcast

59:11SPEAKERknow I'm pretty bearish on generalized,

59:15SPEAKERgenerative machine learning stuff.

59:18SPEAKERThis kind of stuff is really exciting to me

59:19SPEAKER'cause it's doing something that we can't do

59:22SPEAKERwith traditional models.

59:23SPEAKER- Yeah, or this is another one of those examples

59:25SPEAKERof sort of like programmatic use of machine learning, right?

59:28SPEAKER- Yeah.

59:29SPEAKER- They're not like generating art necessarily with this.

59:31SPEAKERThis is more of just kind of a different way of executing

59:34SPEAKERsomething they're already doing in a more efficient way.

59:36SPEAKER- They're taking art that's generated

59:38SPEAKERand making it express more efficiently.

59:41SPEAKER- Yeah, or in this case, not just art, but code, right?

59:44SPEAKERLike shaders are small programs at the end of the day.

59:47SPEAKER- Yeah, but also without changing the artist's intent

59:51SPEAKERis the other thing that's important here.

59:53SPEAKER- Right.

59:54SPEAKER- I talked to a couple of tech artist friends

59:56SPEAKERwho work in this space and like are the people

59:58SPEAKERwho build the materials for a lot of games you've played.

1:00:02SPEAKERAnd generally they were pretty bullish about this.

1:00:04SPEAKERThey were excited about this.

1:00:06SPEAKER- It seems interesting.

1:00:07SPEAKERAnd then in the video presentation at CES,

1:00:09SPEAKERthey flat out were posting like VRAM requirements

1:00:12SPEAKERalongside these things like, hey,

1:00:13SPEAKERthe old version of this takes 28 megabytes

1:00:16SPEAKERand the new one takes 17 or whatever.

1:00:18SPEAKERLike supposedly there are actual

1:00:20SPEAKERtangible memory saving benefits.

1:00:22SPEAKERI'm curious to see where this stuff goes.

1:00:24SPEAKERAnd I don't know how shaders would tie into

1:00:26SPEAKERlike physically based rendering

1:00:28SPEAKER'cause that's the sort of thing you need for ray tracing

1:00:29SPEAKERbecause you need to define the physical properties

1:00:32SPEAKERof all of your materials and surfaces and blah, blah, blah.

1:00:34SPEAKERSo I don't know if this helps there as well,

1:00:36SPEAKERbut like anything that's getting us

1:00:37SPEAKERto a more generalized physical future

1:00:41SPEAKERof the rendering seems good.

1:00:43SPEAKER- Given what the hardware is capable of,

1:00:45SPEAKERmy guess is if we want to have physicality based rendering

1:00:47SPEAKERon real time stuff,

1:00:49SPEAKERaside from relatively limited examples,

1:00:53SPEAKERlike tear down is a good example

1:00:55SPEAKERof they do physicality based stuff,

1:00:57SPEAKERbut it's all hard surfaces and it's not.

1:00:59SPEAKERThere's no, you're not going to make

1:01:02SPEAKERa physicality based renderer right now

1:01:04SPEAKERthat looks like those trolls movies, right?

1:01:06SPEAKERThat's a good example of that.

1:01:08SPEAKER- Yeah, well, we should distinguish here

1:01:09SPEAKERbetween like standard like rendering PBR

1:01:12SPEAKERis just a surface property thing,

1:01:14SPEAKERnot like a physics thing necessarily.

1:01:17SPEAKER- But yeah, this is how we'll get there in real time

1:01:20SPEAKERis my assumption.

1:01:22SPEAKERAnd the word on the street is that Intel, AMD, Qualcomm,

1:01:25SPEAKERMicrosoft, everybody's on board for this.

1:01:27SPEAKERSo we've seen the Nvidia implementation,

1:01:30SPEAKERbut my guess is that we'll,

1:01:32SPEAKERlike as we get into next gen consoles

1:01:33SPEAKERthat have more tensor core, like AI compute capability,

1:01:38SPEAKERthis is where we'll see that stuff start to roll out.

1:01:41SPEAKER- Like maybe, I wonder if this even gets standardized

1:01:43SPEAKERin DirectX maybe.

1:01:45SPEAKER- I would assume so at some point, yeah.

1:01:48SPEAKERNeural textures, same kind of story.

1:01:50SPEAKERIt's using machine learning

1:01:52SPEAKERto do deterministic texture compression

1:01:54SPEAKERat rates that are really ridiculously high.

1:01:58SPEAKER- Oh, interesting.

1:01:59SPEAKERI'm just looking in your notes here.

1:02:00SPEAKERDoesn't decompress into memory?

1:02:03SPEAKER- Yeah, so.

1:02:03SPEAKER- So it's almost like being able to work

1:02:05SPEAKERwith a compressed texture that you don't have to do

1:02:07SPEAKERthat stays compressed as you render it out.

1:02:10SPEAKER- Yeah, that's the exciting thing.

1:02:11SPEAKERAnd the seven to one ratio for compression

1:02:15SPEAKERis the number that they threw out.

1:02:17SPEAKERThat's overexisting texture compression technologies.

1:02:23SPEAKER- Yeah, that seems like a big,

1:02:24SPEAKERI mean, does this work without noticeable artifacts,

1:02:27SPEAKERCBD, but it seems that sounds promising.

1:02:29SPEAKER- Well, I mean, the first step is that it's deterministic,

1:02:32SPEAKERso that means it's the same every time.

1:02:33SPEAKER- Oh, oh, sure, sure.

1:02:35SPEAKER- The phase two is how does it actually look?

1:02:37SPEAKERAnd I haven't spent enough time with it to know that.

1:02:40SPEAKER- Yeah.

1:02:41SPEAKER- The mega geometry is the other thing that's cool.

1:02:43SPEAKERSo one of the things that happens,

1:02:45SPEAKERa lot of objects in games have levels of detail.

1:02:47SPEAKERSo they'll start with a really hero version

1:02:50SPEAKERof your say table or tree or whatever.

1:02:53SPEAKERAnd then you also generate some kind of chunked down,

1:02:57SPEAKERlower fidelity versions of those.

1:03:00SPEAKERAnd depending on the game,

1:03:01SPEAKERyou can go up to like seven, 10, 20 of those sometimes.

1:03:06SPEAKERThose, every time when you're doing

1:03:08SPEAKERtraditional raster rendering, it doesn't matter.

1:03:10SPEAKERIt just pops whatever went in,

1:03:11SPEAKERputs the light map on it, it looks right.

1:03:13SPEAKERWhen you're doing ray tracing,

1:03:14SPEAKERevery time you change the level of detail,

1:03:17SPEAKERyou're ray tracing your pixels for that object,

1:03:19SPEAKERyour rays rather for that object have to be tossed

1:03:23SPEAKERand you have to do that math again.

1:03:25SPEAKERSo it's fairly expensive to change levels of detail

1:03:28SPEAKERin a ray tracing environment.

1:03:29SPEAKERThe mega geometry is a solution for that.

1:03:31SPEAKERIt uses tessellation and a bunch of other stuff

1:03:33SPEAKERto basically make the trees look good.

1:03:38SPEAKER- Yeah.

1:03:39SPEAKERIs this, I may be out of my skis here,

1:03:42SPEAKERbut is this,

1:03:43SPEAKERI looked through some of their kind of developer press docs

1:03:45SPEAKERfor the stuff like, is this somewhat analogous to like,

1:03:48SPEAKERfor example, UV fives nanites,

1:03:49SPEAKERlike virtualized geometry is kind of a thing now,

1:03:52SPEAKERlike it's not just unreal five.

1:03:54SPEAKERI want to say if there's one or two other engines as well,

1:03:57SPEAKERthey're doing this kind of like abstracted geometry

1:03:59SPEAKERthat gets rid of levels of detail and just, you know,

1:04:02SPEAKERthe whole pitch with that thing was like,

1:04:03SPEAKERput your high level hero high geometry assets

1:04:07SPEAKERstraight in the engine.

1:04:08SPEAKERIt just looks the same all the time.

1:04:10SPEAKERPerformance is handled.

1:04:11SPEAKERForget about it.

1:04:12SPEAKERWe're doing it on the back end.

1:04:13SPEAKERBut my understanding is that stuff was pretty CPU heavy

1:04:15SPEAKERbefore now.

1:04:18SPEAKERI wonder, is this a GPU slash,

1:04:20SPEAKERmaybe neural net based solution for that same thing?

1:04:24SPEAKERI don't know.

1:04:24SPEAKER- Well, so I mean,

1:04:26SPEAKERthese are similar solutions for the same problem, right?

1:04:30SPEAKERThe difference is that this works on hardware ray tracing

1:04:33SPEAKERand Lumen and Nanite are for salt.

1:04:36SPEAKERLike Nanite is basically software ray tracing.

1:04:38SPEAKER- Right.

1:04:39SPEAKER- Lumen rather is software ray tracing.

1:04:41SPEAKER- Yeah, and then geometry tessellation and all that stuff.

1:04:44SPEAKER- Yeah, so it should be more performant.

1:04:47SPEAKERIt doesn't run, obviously this is only,

1:04:48SPEAKERthese are all Nvidia implementations for now.

1:04:51SPEAKERSo this is in Alan Wake apparently.

1:04:53SPEAKERLike if you download the latest version of Alan Wake 2,

1:04:57SPEAKERyou can see this in action in the forest scenes.

1:04:59SPEAKER- Yeah, yeah.

1:05:00SPEAKERAs a side note, I have noticed most UE5 games

1:05:03SPEAKERspin the hell out of my CPU fan.

1:05:05SPEAKER- Yeah, UE5 games will spike your CPU pretty good,

1:05:09SPEAKERit turns out.

1:05:11SPEAKERAnd then the last one is the neural faces,

1:05:13SPEAKERwhich is when we get into the,

1:05:14SPEAKERI mean, they talked about ACE,

1:05:17SPEAKERwhich is their character generation

1:05:18SPEAKERwith real natural language stuff, all that.

1:05:21SPEAKER- Yeah, there's like prompt based animation,

1:05:24SPEAKERwhat they were describing as like text to animation

1:05:26SPEAKERwhere the screenshots are like artists typing.

1:05:29SPEAKER- Yeah.

1:05:30SPEAKER- Like waves their arms around

1:05:32SPEAKERand then stretches really tall or whatever,

1:05:34SPEAKERliterally, yeah, we'll see.

1:05:35SPEAKER- That's the kind of stuff that like is real good

1:05:38SPEAKERfor investors and game developers all kind of laugh at.

1:05:41SPEAKER- Yeah, and the neural faces are basically

1:05:43SPEAKERkind of like deep fakes for 3D character models, is that?

1:05:47SPEAKER- Yeah, they basically take the faces generated

1:05:51SPEAKERby the character and a character artist.

1:05:55SPEAKERAnd then they use that to create

1:05:57SPEAKERan upscaled photorealistic version

1:05:59SPEAKERthat they then deep fake onto the model,

1:06:01SPEAKERbasically in real time.

1:06:02SPEAKERAnd it looked real floaty and bad in the videos that I saw.

1:06:06SPEAKERThe Digital Foundry guys saw a slightly newer rev of that.

1:06:11SPEAKERThey talked about it in their post CES breakdown.

1:06:15SPEAKERWhen I talked to character artists about this,

1:06:16SPEAKERthis is where the tech artists were excited

1:06:18SPEAKERabout the neural shaders thing.

1:06:19SPEAKERThe character artists were like, oh, hell no.

1:06:22SPEAKERNo, this is not okay.

1:06:23SPEAKERWe're doing this because we want it to look a specific way.

1:06:27SPEAKERAnd if you look at something like Indiana Jones

1:06:31SPEAKERwhere the character faces are generally very expressive

1:06:34SPEAKERand they replace Harrison Ford's face in Indiana Jones

1:06:38SPEAKERwith a deep faked version of Harrison Ford's face

1:06:41SPEAKERfrom Indiana Jones, it's like, it's gonna look,

1:06:44SPEAKERthis just has real smooth Yoda vibes to me.

1:06:47SPEAKER- Yeah, for sure.

1:06:47SPEAKERI mean, it gets back to the problem of determinism again,

1:06:50SPEAKERright, because like a pattern on a quilt,

1:06:51SPEAKERyou know exactly, that's supposed to look the same

1:06:53SPEAKERevery time and you can massage the model

1:06:55SPEAKERuntil you're getting the pattern on the quilt that you want.

1:06:57SPEAKERLike the human's facial expressions are very

1:07:02SPEAKERnot understandable by a computer, let's say.

1:07:04SPEAKERSo like getting what you want every time out of this,

1:07:07SPEAKERthat seems way trickier and maybe impossible.

1:07:10SPEAKER- Also it's unnecessary, right?

1:07:11SPEAKERLike the point is you're gonna,

1:07:14SPEAKERlike if you're building a game, you're choosing,

1:07:16SPEAKERand you're thoughtful about it, you're making a good game,

1:07:18SPEAKERyou have made style choices and artistic choices

1:07:22SPEAKERbased on A, the capability of your team

1:07:25SPEAKERand the size of your team and the size of your art staff,

1:07:28SPEAKERbut also the game that you wanna make.

1:07:30SPEAKERAnd if you take a game that has like a real defined style

1:07:33SPEAKERand then deep fake on some gen AI faces,

1:07:37SPEAKERit's gonna look like crap and it's gonna, it's just like,

1:07:42SPEAKERthis again, a lot of times, a lot of the AI stuff

1:07:46SPEAKERand the places where the AI stuff has had the biggest gap

1:07:49SPEAKERbetween what people who make AI software want,

1:07:54SPEAKERthink I want and what I actually want

1:07:56SPEAKERis when they're taking away artist intent

1:07:59SPEAKERand replacing it with a machine.

1:08:01SPEAKERAnd this is that, like this is, I was bummed to see this.

1:08:04SPEAKER- Yeah.

1:08:05SPEAKER- So anyway, that's the kind of,

1:08:09SPEAKERI wanna hit one more thing

1:08:11SPEAKERthat's kind of off of our, well, two things.

1:08:13SPEAKEROne is that these cards are high power.

1:08:17SPEAKERThe min spec on the 5090 is a thousand watt power supply.

1:08:21SPEAKER- Yikes, min?

1:08:22SPEAKER- Min spec, it uses the 12 volt two by six,

1:08:25SPEAKERwhich is the like the rev 2.1 12 volt high power.

1:08:30SPEAKER- Like that's the 12V HPWR to be clear.

1:08:33SPEAKER- Yeah, it's--

1:08:35SPEAKER- Like that's the ATX, the point is that's the ATX one,

1:08:37SPEAKERright, like that's the one that you should have

1:08:39SPEAKERif you have an ATX 3.0 power supply.

1:08:41SPEAKER- Yeah, and it's physically compatible

1:08:45SPEAKERwith the 12 volt high power.

1:08:47SPEAKERMy 12 volt high power that was in my power supply

1:08:51SPEAKERhad a 450 watts number written on the side of it.

1:08:54SPEAKER- Oh, interesting.

1:08:55SPEAKERMine might have that as well, actually.

1:08:57SPEAKERI need to look.

1:08:58SPEAKER- So I had to upgrade power,

1:08:59SPEAKERI mean, I had an 850 watt power supply,

1:09:00SPEAKERso I had to upgrade anyway.

1:09:02SPEAKER- Mine might be 600 actually.

1:09:04SPEAKER- But yeah--

1:09:05SPEAKER- I have a 1200 watt.

1:09:06SPEAKER- The 1200 watt supply had a 600 watt one.

1:09:09SPEAKER- Okay.

1:09:10SPEAKER- From our, I mean, like when we did power testing

1:09:13SPEAKERat PC World, we were seeing total,

1:09:17SPEAKERwe didn't break out the power supply,

1:09:19SPEAKERthe power draw from the video card,

1:09:22SPEAKERjust 'cause that's fraud and you can't isolate

1:09:25SPEAKERthe slot power and all the other things.

1:09:27SPEAKERWe were seeing total system power

1:09:30SPEAKERthat was significantly higher with this new card

1:09:32SPEAKERthan the older cards, and I'm trying to find

1:09:35SPEAKERthe numbers here, but--

1:09:37SPEAKER- I think I talked to Adam briefly,

1:09:39SPEAKERand I think it was like over 700 watts total draw.

1:09:42SPEAKER- Yeah, that seems right.

1:09:44SPEAKERThe Cyberpunk, Cyberpunk average power consumption

1:09:47SPEAKERwas 533 watts with spikes at like 702 watts.

1:09:51SPEAKERAgain, this is for a 7950 X3D and a 64 gig system, I think.

1:09:57SPEAKER- Okay, maybe that's maybe an important clarification.

1:09:59SPEAKERThose were 700 watt peaks not sustained.

1:10:02SPEAKER- Yeah, the average for the Cyberpunk run was 533 watts.

1:10:06SPEAKER- Okay.

1:10:07SPEAKER- That includes a little bit of idle on both ends

1:10:09SPEAKERas it's kind of ramping and loading and all that stuff.

1:10:11SPEAKER- All right, good to know.

1:10:12SPEAKER- The other thing, the other thing is that the packaging,

1:10:17SPEAKERlike they, like I wanna shout this out

1:10:20SPEAKERbecause they don't have to do this,

1:10:21SPEAKERand it's like I think the 4090 packaging,

1:10:24SPEAKERespecially for the Founders Edition, was kind of overkill

1:10:27SPEAKERand had a bunch of plastic and crap in it,

1:10:29SPEAKERand it made this great big cool presentation box

1:10:32SPEAKERfor a thing that you're gonna,

1:10:33SPEAKERprobably most people are gonna recycle

1:10:35SPEAKERabout five minutes after they put the card in their system.

1:10:38SPEAKERThe 5090 box is all paper.

1:10:41SPEAKER- Okay.

1:10:42SPEAKER- It doesn't use any plastic fasteners even

1:10:44SPEAKERto seal the internal shipping material and all that,

1:10:48SPEAKERso it's all, it looks like a dog bone,

1:10:51SPEAKERand in the cutoutsides, you slide these little loops in

1:10:55SPEAKERthat slide into notches and they hold it closed.

1:10:58SPEAKERIt doesn't have any plastic to throw away at all in there

1:11:01SPEAKERthat I saw except for maybe the adapters,

1:11:03SPEAKERwhich you may or may not use.

1:11:05SPEAKER- Yeah.

1:11:06SPEAKER- And I just wanted to shout that it's a much smaller box

1:11:08SPEAKERthan the 4090, which means they can ship more of them

1:11:10SPEAKERin a crate, which means less carbon, so that's good.

1:11:13SPEAKERLike, Sony made a very notable or profound move

1:11:18SPEAKERtoward that kind of sustainable packaging,

1:11:19SPEAKERstarting with the regular PS5.

1:11:21SPEAKERI don't think they did anything before that,

1:11:23SPEAKERbut anyway, like, you know, I got the PS5 Pro recently,

1:11:25SPEAKERand like, they literally use like cardboard twist ties now.

1:11:30SPEAKER- Yeah.

1:11:30SPEAKER- Like, obviously there's still a metal wire

1:11:32SPEAKERin the twist ties, but the outer covering is cardboard

1:11:34SPEAKERrather than, or paper rather than plastic.

1:11:37SPEAKERLike, everything in that box is biodegradable now,

1:11:40SPEAKERwhich is kind of nice to see.

1:11:41SPEAKER- Yeah, the whole box can go in the recycling bin,

1:11:43SPEAKERor I guess probably even the compost bin,

1:11:45SPEAKERI don't know what the ink situation is.

1:11:47SPEAKERBut yeah, I was impressed, and it was a nice change.

1:11:51SPEAKER- Yeah.

1:11:53SPEAKER- And like, there's no like plastic shock material

1:11:55SPEAKERor anything like that, it's all just coordinated cardboard

1:11:57SPEAKERall the way down, so props to them.

1:11:59SPEAKER- Yeah.

1:12:00SPEAKER- And then we gotta talk about my real problem

1:12:02SPEAKERwith the 5090.

1:12:03SPEAKER- Go on.

1:12:04SPEAKER- Well, I've had three slot cards in my machine

1:12:05SPEAKERfor a while now, Brad.

1:12:06SPEAKER- All right.

1:12:07SPEAKER- And when I had to put this in,

1:12:09SPEAKERI couldn't remember where I put the slot cover

1:12:11SPEAKERfor the third slot, 'cause this is a two-slot card.

1:12:14SPEAKER- Wait, wait.

1:12:15SPEAKER- On the back of the computer?

1:12:16SPEAKER- Oh, oh, oh, you mean like the bracket?

1:12:18SPEAKER- Yeah.

1:12:19SPEAKER- The bracket that covers the, eh.

1:12:21SPEAKER- Yeah, like it took me forever, it was a nightmare.

1:12:24SPEAKERIt was awful.

1:12:25SPEAKER- You know, it was probably a good problem to have.

1:12:26SPEAKER- I mean, it was okay, it was an okay problem to have.

1:12:29SPEAKERBut yeah, it's weird having a two-slot card in there again.

1:12:32SPEAKER- Yeah, yeah, I believe it.

1:12:33SPEAKERI also am ready for, I need a smaller GPU as well,

1:12:36SPEAKER'cause this roughly four-slot card

1:12:39SPEAKERis making motherboard choice kind of difficult

1:12:41SPEAKERwhen you need to use the other, you see how I express slots?

1:12:43SPEAKER- Yep, yep.

1:12:44SPEAKER- And I'm never just, or more general,

1:12:46SPEAKERI'm just ready for maybe a move

1:12:47SPEAKERto smaller graphics cards in general,

1:12:50SPEAKERhopefully we'll get there, but anyway.

1:12:52SPEAKER- You should just stop playing games and, you know,

1:12:54SPEAKERjust get a passive, and just use it.

1:12:57SPEAKERYou can use the integrated, get an AMD machine,

1:12:59SPEAKERjust use that built-in CPU, built-in GPU.

1:13:01SPEAKER- This Raptor Lake has got a fine iGPU in it.

1:13:04SPEAKERI'll just go back to that.

1:13:05SPEAKER- There you go, you're done, perfect.

1:13:06SPEAKER- That'll run Fortnite, right?

1:13:07SPEAKER- Yeah.

1:13:09SPEAKER- No, probably not.

1:13:10SPEAKER- It'll run, I bet it'll run the Switch renderer.

1:13:13SPEAKERYou can switch to the Switch renderer if you're on--

1:13:15SPEAKER- You can?

1:13:16SPEAKER- Yeah. - On PC?

1:13:16SPEAKER- Uh-huh, they don't call it the Switch renderer,

1:13:18SPEAKERthey call it performance mode,

1:13:20SPEAKERbut it's basically like a DX11 looks like,

1:13:22SPEAKERit looks just like the Switch when you play it.

1:13:24SPEAKER- That's funny, I shouldn't, man,

1:13:25SPEAKERI wish they did call it the Switch renderer,

1:13:27SPEAKERbut I get why they don't.

1:13:28SPEAKERAnyway, new graphics cards.

1:13:30SPEAKER- Yeah, so if you have questions,

1:13:32SPEAKERthis should have been a question week this week,

1:13:34SPEAKERbut we will do questions next week, as is our way.

1:13:38SPEAKERSo if you have Q's, you can send them to TechPod

1:13:40SPEAKERat content.town and we will turn them to A's

1:13:43SPEAKERthis time next week.

1:13:45SPEAKERAnd I guess we've reached the part of the show

1:13:47SPEAKERwhere we thank our patrons.

1:13:49SPEAKER- We have, thank you, patrons.

1:13:50SPEAKER- Thank you, patrons.

1:13:52SPEAKERAs always, "Brad and Will Made a TechPod"

1:13:54SPEAKERis a listener-supported show,

1:13:55SPEAKERso without you, we would not be here,

1:13:58SPEAKERwhich means if you want to find out how to support the show,

1:14:00SPEAKERyou can pay us five bucks, it's at patreon.com/techpod.

1:14:05SPEAKERFor that $5, you get two, well,

1:14:08SPEAKERI guess it depends how you count them,

1:14:09SPEAKERyou get two things, but they both contain multitudes.

1:14:13SPEAKEROne of the things is access to the TechPod Discord,

1:14:16SPEAKERwhich is full of beautiful nerds

1:14:17SPEAKERwho talk about all sorts of fascinating stuff all day.

1:14:21SPEAKERI dip into different channels

1:14:22SPEAKERand I learn something pretty much every time.

1:14:24SPEAKER- It's a fount of wisdom in all respects.

1:14:27SPEAKERLet's say, okay, should not even get into this,

1:14:30SPEAKERthis is patron episode material.

1:14:32SPEAKER- Uh-oh.

1:14:33SPEAKER- Let's say you have become bent on self-signing

1:14:35SPEAKERall of your EFI binaries,

1:14:37SPEAKERso you can turn on secure boot on all of your computers

1:14:40SPEAKERwhen you multi-boot a bunch of esoteric operating systems

1:14:43SPEAKERand you need to know about that stuff.

1:14:45SPEAKER- Okay, I have questions,

1:14:47SPEAKERbut I'm gonna hold them for the patron episode.

1:14:49SPEAKERDoes somebody know how to do that?

1:14:50SPEAKER- Oh yeah, you absolutely,

1:14:52SPEAKERyou can literally open SSL your own,

1:14:54SPEAKERokay, caveat to what I'm saying here.

1:14:57SPEAKERMy current understanding is this is possible.

1:15:00SPEAKERYou can generate your own keys,

1:15:02SPEAKERsign all the EFI binaries,

1:15:03SPEAKERwhich are the things that those are the,

1:15:05SPEAKERit's basically the boot loaders

1:15:06SPEAKERthat launch your operating systems.

1:15:08SPEAKERYou can sign all those yourself

1:15:09SPEAKERand then enroll your public keys into your UEFI

1:15:13SPEAKERas valid secure boot keys.

1:15:15SPEAKERAnd then you can secure boot basically anything

1:15:17SPEAKERas an EFI binary.

1:15:19SPEAKER- But then what happens if you,

1:15:21SPEAKERyou just have to make sure you back up your-

1:15:22SPEAKER- I don't know.

1:15:23SPEAKER- Keys before you update your BIOS.

1:15:26SPEAKER- I don't know yet.

1:15:27SPEAKERWell, I mean, worst case you just turn secure boot back off

1:15:29SPEAKERand everything, but it's just fine.

1:15:31SPEAKER- When I had this happen and I turned secure boot back off,

1:15:33SPEAKERWindows just never booted again.

1:15:34SPEAKERI had to reinstall Windows.

1:15:36SPEAKER- Yeah, I think that's changed now

1:15:38SPEAKER'cause I did turn it off the other day, original.

1:15:40SPEAKERI think now you have to have secure boot on to install 11,

1:15:43SPEAKERbut then you can turn it off afterwards is the case.

1:15:46SPEAKER- You can have secure boot off.

1:15:47SPEAKERYou just have to, you have to have a TPM2 module.

1:15:50SPEAKER- Yes, I guess maybe that's the case actually.

1:15:51SPEAKER- Yeah, you don't, secure boot's optional.

1:15:54SPEAKER- It's a whole rabbit hole.

1:15:55SPEAKERWe'll get into it at some point.

1:15:57SPEAKER- Oh, the other thing you get that contains multitudes

1:15:59SPEAKERis the patron episode where we talk about stuff

1:16:01SPEAKERlike Brad's adventures with secure boot

1:16:04SPEAKERand self-signing certificates and let's encrypting, I guess.

1:16:09SPEAKERSo to do that, you go to patreon.com/techpod.

1:16:12SPEAKERAgain, it's patreon.com/techpod

1:16:15SPEAKERand we appreciate the support from everybody.

1:16:18SPEAKERIf you don't have five bucks to do that, that's fine.

1:16:19SPEAKERJust leave a review someplace like iTunes or I don't know,

1:16:23SPEAKERSpotify does reviews now or Deezer, Deezer has reviews.

1:16:26SPEAKERWe love Deezer.

1:16:27SPEAKER- That's not a real platform.

1:16:28SPEAKER- Deezer is a real platform.

1:16:29SPEAKERWe are on Deezer, right?

1:16:30SPEAKERYou can go to deezer.com

1:16:32SPEAKERand look for Brad Wilmer to talk about

1:16:33SPEAKERand we are there, Brad.

1:16:34SPEAKER- Whatever you say, man.

1:16:36SPEAKER- But we've reached the part of the show

1:16:37SPEAKERwhere we thank our patrons.

1:16:39SPEAKERStarting since the end of the month

1:16:41SPEAKERwith our executive producer tier patrons,

1:16:44SPEAKERincluding Andrew Slosky, Bunny Thing, David Allen,

1:16:47SPEAKERJames Kamik, Jordan Lippett, Twinkle Twinkie,

1:16:51SPEAKERand Pantheon, makers of the HS3 high-speed 3D printer.

1:16:55SPEAKERSince this is the end of the month,

1:16:55SPEAKERwe also thank our associate producer tier patrons,

1:16:58SPEAKERincluding Alejandra Navarro, Andre M. Burke, P.E.,

1:17:03SPEAKERor it's Andre M. Burke, P.E.

1:17:06SPEAKERI don't know what P.E. stands for.

1:17:08SPEAKERAndrew Dicey Scholdyce, Arthur Gies, Ben Tallman,

1:17:12SPEAKERBrutal Kerfuffle, Eric Klein, Felix Kramer,

1:17:16SPEAKERGraham Banks, Jad Rita, Just Associate Wedge,

1:17:20SPEAKERMatt Walker, parentheses Walkman8080, close parentheses,

1:17:24SPEAKERMike Etheridge, Nathan Phelps, Sanchik Kumar,

1:17:27SPEAKERSteve Lin, Thomas Shea, and Tom Hilton.

1:17:30SPEAKERThanks everybody so much.

1:17:31SPEAKERWe appreciate the support.

1:17:33SPEAKER- Yeah, we sure do.

1:17:34SPEAKER- And shout out to the people who put

1:17:35SPEAKERthe phonetic pronunciation of their names

1:17:37SPEAKERand their names when it's tricky,

1:17:39SPEAKERbecause that really helps a lot.

1:17:41SPEAKER- Yeah, sounds like an example

1:17:42SPEAKERof some secretly encoded information,

1:17:44SPEAKERbut we'll get into that.

1:17:45SPEAKER- Oh, secretly encoded information.

1:17:46SPEAKERIf you have that, we have a form that you can submit ideas

1:17:50SPEAKERfor secretly encoded information in,

1:17:52SPEAKERwhich I guess we could put in the show notes.

1:17:54SPEAKER- I will put that in the show notes.

1:17:55SPEAKER- Perfect.

1:17:56SPEAKERAnd Brad, this has been a pleasure.

1:17:58SPEAKERI guess we'll talk to you next week about questions.

1:18:01SPEAKERTechPod at contact.town.

1:18:03SPEAKERSee you all next week.

1:18:04SPEAKERConsider the environment before printing this podcast.

1:18:07SPEAKERBye everybody.