0:00

In this episode of the Full

0:02

Nerd, we're talking about AMD's Radeon

0:04

Delay, the take from NVIDIA's Editor's

0:06

Day, and what reviewers of Intel's

0:08

B570 have to say, and we're

0:10

going to turn your Q's into

0:12

A's. Welcome

0:23

to episode 332 of the Full

0:26

Nerd, the best weekly podcast about

0:28

PC building news. I'm your fill

0:30

-in -fill -in host, Will Smith, joined

0:32

today by co -host Brad Charcus. He's

0:35

my permanent co -host, Brad Charcus. Hello,

0:37

internet. Elena Yee. Hello.

0:40

And on the verticals

0:42

and horizontals, Alex Estevez. Hello,

0:45

hello. And walking behind us, just

0:47

in the shot, Adam Patrick Murray.

0:49

He's, get back to benchmarking monkey.

0:52

Get in there. Crank those bits. Not

0:54

that we're doing that. I don't know if we're

0:56

allowed to talk about that. But welcome. Hello, everybody.

0:58

Hi. Hello. Hi. Hey, how's it

1:00

going? This is great. This Tuesday feels like a Monday, but

1:02

in a good way. Well,

1:04

yesterday was a Monday, but

1:06

it was a Monday off. Yeah. I

1:09

forgot because I, well, I was unboxing things,

1:11

I guess is what I'm allowed to say.

1:13

But yeah, so what did you guys do

1:15

on your day off? Anything fun? I

1:19

did the opposite of fun. I bought a car after my

1:21

car was t -boned last month. I walked into the place

1:23

knowing exactly which one I wanted, took it for a

1:25

test drive, a couple other ones while I was there. The

1:28

whole process still, still took over seven hours.

1:32

Today to do some more bank

1:34

paperwork and stuff. So it's not fun.

1:36

Try not to buy a car. If you don't have to, they're

1:38

a lot more expensive now too. Oh

1:40

yeah. How bad is it? I

1:43

don't want to SUV. I only

1:45

like cars like sedans. I was looking

1:47

for used in my area. There

1:50

were only two used cars available

1:52

with under 100 ,000 miles

1:54

and they cost 15 ,000

1:56

plus. Wow. But

2:00

you got what you wanted? I did. I

2:02

ended up with a pretty good car at

2:04

a pretty good deal considering the you know

2:06

market so I'm happy with it and I

2:08

have been stuck in my house for a

2:11

month so I'm very happy to have the

2:13

opportunity go get my haircut. I was just

2:15

saying what was the first thing you did

2:17

you went out you clearly haven't gotten your

2:19

hair cut yet but that's coming. Yeah. I'm

2:21

in the same boat. It's a duck and

2:23

a crowd after a grueling weekend. Sorry,

2:26

look, you spent seven hours at the

2:28

car dealership yesterday. You could have at

2:30

least swung by the barbers. No, no,

2:32

it's fine. I went on the highway

2:34

where it gets up to 75 miles

2:36

per hour up in the North country

2:38

and just drank the music and saw

2:40

how it rode and I liked

2:42

it. Nice, that's awesome. Congratulations,

2:45

I'm glad you got the car. Elena,

2:47

anything exciting on your end?

2:49

No, I woke up and forgot that there

2:51

was a holiday. And then I, it's like, oh

2:53

wait, I can go back to sleep, okay. Because I am

2:55

so jet-lagged, like still, like, I don't want to even jet-lagged,

2:57

it's like, I was more of a night owl before I

3:00

went to Europe, and I have returned as a morning person,

3:02

and emotionally and intellectually, I don't know what to make of

3:04

this. So now the sun comes up at 7 AM, my

3:06

body's like, cool. I'm like, what's like, what's happening, what's happening,

3:08

what, what, what, what, what, what, what, what, what, what, what,

3:10

what, what, what, what, what, what, what, what, what, what, what,

3:13

what, what, what, what, what, what, what, what, what, what, what,

3:15

what, what, what, what, what, what, what, what, what, what, what,

3:17

what, what, what, what, what, what, what, what, what This

3:19

is the worst part of getting old for

3:21

me is the is the now I go

3:24

to bed at like midnight and I wake

3:26

up at 6 or 7 530 sometimes and

3:28

I'm like well I'm not going back

3:30

to sleep now so I just get

3:32

up and do stuff it's awful

3:34

it's terrible productivity yeah I

3:36

know should we should we

3:39

talk about the news the

3:41

computer news the the hard

3:43

PC hardware news Oh, there's computer

3:45

news happening? Well, we were taking

3:47

a three-day weekend? Sure. I mean,

3:49

look, the Friday news dump was

3:51

flush on Friday. I don't know

3:53

if you, like, people were getting

3:55

all the bad stuff out there. But,

3:57

yeah, the top news, radion

3:59

delay. Originally at CS I think

4:02

said that they were going to

4:04

talk about next gen RDA for

4:06

stuff by the end of January,

4:08

I think was the, I don't

4:10

know if they actually gave a

4:12

date, but it seemed like they

4:14

were implying that they were talking

4:16

about in the few weeks. They

4:18

didn't give a hard date. The

4:20

fact that it was in the

4:22

announcement that got sent out to

4:24

press just the day before, followed

4:26

by there being a bunch of

4:28

models for sale on the show

4:30

floor. Made people think it was

4:32

going to be very soon. And

4:34

Frank Azor, when he was asked

4:36

in a roundtable, said, we're talking

4:38

about weeks before the graphics card

4:40

show up. And he said, when

4:42

we say quarter one, you know,

4:45

we're not talking about March 31st.

4:47

But according to this new statement

4:49

from David McAfee, who's in charge

4:51

of client computing, I think, over

4:53

at AMD. 90-70 series will be

4:55

in March now. So maybe not

4:57

March 31st, it sounds like, but

4:59

sometime in March. Wow. I mean,

5:01

does this, so can we read

5:03

anything into this? Like what, I

5:05

got the feeling that they were

5:07

kind of waiting and seeing what

5:09

invidious rolled out with was why

5:11

they didn't announce stuff at CES

5:13

necessarily. The 45-minute press conference maybe

5:15

was a happy, happy excuse. But

5:17

I'm curious if like there. They're

5:19

not able to make any appreciable

5:21

board design changes at this point.

5:23

So it's got to be just

5:25

adjusting pricing and positioning, I guess.

5:27

My gut feel around all this

5:29

is they expected the 5070 and

5:31

5070 TI to be a lot

5:33

more money, especially because they just

5:36

repositioned this generation. They changed their

5:38

branding to call it a 9070

5:40

to be a 5070 competitor. So

5:42

my gut says they were hoping

5:44

to price it a little bit

5:46

higher than they might be able

5:48

to now. And this

5:50

will give them time to work on

5:52

DLS or FSR4. Yeah. Because, you know,

5:55

with in video launching 75 games, DLS4

5:57

coming out this month, next month. that

5:59

stuff. This will give them time to

6:02

actually have episode foreign titles, hopefully, so

6:04

hopefully that's some of that as well, but

6:06

I think they just really want to see

6:08

what the 50-70 brings to the table. Because

6:10

invidious pent all of its time talking

6:13

about multi-frame generation and RT

6:15

and stuffs, you can't get a real

6:17

clear picture of where they expect things

6:19

to be. Yeah, I mean, the most

6:21

we could tell from those benchmarks, we'll

6:23

talk about this a little bit in

6:26

a little bit in a minute, obviously.

6:28

is like looks like maybe 20 or

6:30

30% uplift gen gen over gen on

6:32

raster stuff and then like how how

6:34

the MFG stuff actually works in practice

6:37

is the is it's kind of the

6:39

billion dollar question right now from where

6:41

I'm sitting so yeah so and MD

6:43

I I wonder if they were caught

6:45

a little bit surprised by

6:48

multi-frame generation and how

6:50

heavily invidia focused

6:52

on multi-frame generation

6:54

on multi-frame generation.

6:56

And it's just waiting to see. Well,

6:58

there we go. I'd rather a lot of people

7:00

are bummed. I would rather them wait and

7:03

see and figure out where the right

7:05

position for this product in the

7:07

marketplace is and launch in that spot. Like,

7:09

I'm a big believer. And I think it

7:11

was, uh, me, Motu said you can only,

7:13

you know, everyone, nobody remembers that

7:16

a game was delayed. They just remember

7:18

if it came out bad. I think

7:20

that AMD's suffered from that in the

7:22

past few generations. So I'm totally fine

7:25

with them taking the next month or

7:27

two to try to nail the right

7:29

spot. It would be nice to have a

7:31

hardware launch, just be smooth.

7:33

Just one, just one. Yeah, for a lot,

7:35

well. Yeah, look, it's been a long

7:38

time, Elena, since we had something

7:40

that just came out and was

7:42

like, oh, hey, this is what

7:44

we expected. It's working as we

7:46

expected. Everything's good. No problemos.

7:48

So, I don't know. I have a question.

7:50

Yeah, invidious. We don't know what their what

7:52

their launch is looking like. Who knows? We'll

7:54

we'll find out in a I don't know

7:57

what embargoes are up whenever that is. I

7:59

have a question. So at this for a

8:01

perspective of somebody who might want a

8:03

card. Does the extra time also give

8:05

them the ability to kind of like

8:07

build up their stock as well so

8:09

that not necessarily doesn't feel like a

8:12

paper launch? Presumably yeah. Yeah. Yeah. And

8:14

right now there's rumors. Go ahead Brad.

8:16

Right now there's rumors of that Blackwell

8:18

the 50 series will be kind of

8:20

short immediately at launch because it's launched

8:22

in January with the Chinese New Year

8:25

coming up. So luncheon in March would

8:27

give AMD time for everyone to get

8:29

back to those factories in Asia and

8:31

work get a lot more of these

8:33

things out. Well, and and also presumably

8:35

it means that the AIB vendors will

8:38

have time to get their parts lined

8:40

up, which is obviously much more important

8:42

for AMD than it is for invidious

8:44

these days, right? I feel like the

8:46

AIB's thought it was going to launch

8:48

a lot sooner than this. I can't

8:51

say much more than that. some of

8:53

them expected as well. That's interesting. Yeah,

8:55

so, okay, so anything else on the

8:57

aimed on the radion delay, I feel

8:59

like that's pretty much the high stuff,

9:02

right? Yeah, that's it. It just kind

9:04

of feels like the ultimate, you know,

9:06

Lucy pulling the football on Charlie Brown

9:08

again. It was like, wow, really, just

9:10

don't even mention it. I mean, kind

9:12

of, I kind of, yeah, but also

9:15

it kind of feels like maybe, like,

9:17

like they didn't really announce anything at

9:19

CES. Like we had a conversation before

9:21

the show started about whether we should

9:23

even call this a delay because they

9:25

were so vague about what was coming

9:28

at CES that we weren't sure what

9:30

they said we had to go look

9:32

it up. Yeah I don't think it's

9:34

an actual I wouldn't count this as

9:36

a delay but it's delayed compared to

9:38

what everyone's expected. Yeah I mean this

9:41

is the benefit of going second is

9:43

you can you can change your game

9:45

plan and now they can price according

9:47

to how these 50-70s actually perform versus

9:49

there would have been a nightmares conversation

9:52

backstage where you're like, okay, if the,

9:54

if the, if the, if the, if

9:56

the, if the, if the, if numbers

9:58

are real and people are actually going

10:00

to use this we have to price

10:02

this at this and if the MFT

10:05

numbers are fake and we're not going

10:07

to do this then we have to

10:09

price this at this and they're going

10:11

to make a guess and they're going

10:13

to base the next two 18 months

10:15

two years of graphics card sales on

10:18

guesses based on rumors no this is

10:20

the right thing to do you see

10:22

to argue about for another three months

10:24

no This is what the redic craves.

10:26

I'm also wondering, I'm also wondering if

10:29

this is a good sign, right? Because

10:31

I feel like a lot of times

10:33

when you're not confident in your product,

10:35

you will rush out to be the

10:37

first just to get it out of

10:39

the way so that you don't have

10:42

a negative comparison to what follows or

10:44

like to what you're following. And so

10:46

the fact that they're holding on to

10:48

it to me kind of suggests that

10:50

I think they have confidence in this

10:52

product. Maybe it could all like the

10:55

other thing I wondered is if the

10:57

Intel stuff has thrown them for the

10:59

poor loop Maybe they weren't expecting competition

11:01

at that at that low end and

11:03

as the company now that has come

11:05

out and said hey we're talking to

11:08

the low and mid range Like hey

11:10

we've lost the high-end mid-range and we've

11:12

lost the low-end is a real that's

11:14

a bad story to have to tell

11:16

from a from a sales perspective so

11:19

I think it's Maybe I'm a glass

11:21

half empty kind of guy, but like

11:23

Elena is the voice of optimism and

11:25

Brad and I are over here on

11:27

the grumpy old guy side. There's a

11:29

reason we have all this white in

11:32

our beards. Yeah, maybe that's why. Like

11:34

once I get the beard and I

11:36

like... Try to like blend in maybe

11:38

it'll be different I saw a real

11:40

good one on the tick-tock shop the

11:42

other day I should have just mashed

11:45

the button for you later Let's talk

11:47

about invidious editors day stuff because we

11:49

can the embargo lifted on that a

11:51

couple days ago We've we Adam was

11:53

there on site Brad and I have

11:55

had access to materials and stuff for

11:58

this a long time. I don't know

12:00

about you, but I spent a lot

12:02

of time reading white papers this weekend

12:04

because I was wildly curious about what

12:06

they've done with this architecture. And they

12:09

broke down the blackwall stuff, which kind

12:11

of already talked about last week a

12:13

little bit, so we don't need to

12:15

get into a ton of detail on

12:17

like skews and all that stuff. But

12:19

they talked about the changes of the

12:21

tensor cores, they talked about DLS4, they

12:24

talked a lot about multi-frame generation, which

12:26

is the thing I think I think

12:28

I'm most curious about. And then they

12:30

talked about some like crazy

12:32

stuff from the future. So

12:35

I guess we start with

12:37

Blackwell, I guess is the

12:40

place to begin. Blackwell, the

12:42

new generation, GPU

12:44

core, new tensor cores,

12:47

which we weren't exactly

12:49

expecting, a new version

12:52

of DLS. Upgraded rate tracing

12:54

cores. And a new, the

12:56

board design. I feel like some of

12:59

the stuff we should just save at

13:01

this point for reviews, because the board

13:03

design is, is, we talked about it

13:05

a lot last week, I don't think

13:07

we need to get into that. Is

13:09

there anything on the architecture stuff that

13:11

you got out at Editor's Day, materials

13:13

that you didn't have from Jensen's talk

13:15

and from what we talked about last

13:17

week? A couple interesting high-level thoughts.

13:19

I don't want to dig far

13:21

into it, because it's a lot

13:23

of information to process. I find it

13:25

interesting that they moved back to

13:27

a shared pipeline for FP and

13:29

I&T calculations to make it, because

13:32

the past, that's what they used

13:34

to have, the past two generations, they've

13:36

had separate pipelines for integer

13:38

and F floating point performance

13:40

tasks, this time they're putting

13:42

it all back together in one

13:44

big shared one with the switching. The

13:47

amount like that they're putting into

13:49

these tensor cores this generation like

13:52

really surprises me to me I

13:54

think the defining aspect of this

13:57

generation will be DLS for I

13:59

think there's reason they're pushing for

14:01

the the driver things you can

14:03

I forget what they call it

14:05

so you can switch out older

14:08

DLS versions and force oh DLS

14:10

for yeah the forced the forced

14:12

DLS for in the invady app

14:14

yeah yeah yeah I think there's

14:16

a reason to do that come

14:18

out with 75 right out of

14:20

the gate they've really overhauled those

14:22

tensor cores I think the new

14:25

FP4 capabilities in these tensor cores

14:27

is going to be what drives

14:29

a lot of that new DLS

14:31

stuff. And I think if we

14:33

do wind up seeing it back

14:35

ported to the 40 and 30

14:37

series, 20 series generations, it won't

14:39

perform as well because they have

14:42

an outright set it. But I

14:44

was watching an excellent digital foundry's

14:46

interview with Brian Kata Rezano. I

14:48

hope I got your name right

14:50

Brian. The DLS guy, AI guy

14:52

at Invidia. And he strongly implied

14:54

that, you know, they designed the

14:57

Blackwell architecture in hand to work

14:59

with DLS for implying that it's

15:01

leaning on that FP4. So that's

15:03

interesting to me. The other interesting

15:05

bit is that it's really just

15:07

kind of Ada lovely sounds terroids.

15:09

Like... like the raster pipeline you

15:11

mean yeah it's the same the

15:14

raster pipeline's different but it's still

15:16

the same TSMC node like yeah

15:18

so I'm curious to see where

15:20

it lands in reviews yeah the

15:22

the memory bus is a little

15:24

bit wider but we I think

15:26

we knew that from the from

15:29

the Jensen talk yeah like for

15:31

me for me the exciting thing

15:33

is the DLS4 stuff because it's

15:35

it's like The thing that Jensen

15:37

said on stage that seemed like

15:39

it was ludicrous when we're hearing

15:41

it on stage was, hey, we're

15:43

going to take a 1080P frame,

15:46

render that once, and then generate

15:48

four 4K frames off of that.

15:50

The first one using super sampling,

15:52

and then the other three using

15:54

the AI optical warp multi-frame. generation

15:56

tech? Yeah, reflex. Well, reflex, so

15:58

reflex one for that. Reflex two

16:01

is separate. Okay. That was unclear.

16:03

That I didn't get until I

16:05

was like that took a lot

16:07

of digging to figure out. The

16:09

reflex two stuff is only in

16:11

two games that they've announced so

16:13

far, the finals and valorant. It's

16:15

not out yet, and it's the

16:18

thing that reduces the latency from

16:20

75%. They're not using that for

16:22

the multi-frame generation. The thing that

16:24

was interesting is they said they

16:26

can do the, in the white

16:28

papers, they said they can do

16:30

the multi-frame generation stuff on one.

16:32

tensor cycle or one tensor call

16:35

for three frames. So they're taking

16:37

that 1080P image and generating three

16:39

frames off of that and then

16:41

optical warping them to match input.

16:43

So they feel lower latency even

16:45

though I guess they're not. It

16:47

seemed like they did a lot

16:50

of tensor core stuff to make

16:52

those frame generations happen faster and

16:54

later in the pipeline so you

16:56

can account for more stuff as

16:58

it's happening. Anyway. Yes, I think

17:00

truly think like DLS4 if it's

17:02

also or not is going to

17:04

be the defining moment of this

17:07

generation. Yeah, like because the thing

17:09

is we haven't like we have

17:11

cards where I think we're allowed

17:13

to say we have cards because

17:15

we were we were post on

17:17

boxings yesterday. We're not allowed to

17:19

talk about anything beyond that. The,

17:22

if the DLS4 stuff doesn't feel

17:24

good, none of this works, right?

17:26

Yeah. And because like the uplift

17:28

generation over generation from what they

17:30

showed at Editor's Day is like

17:32

30% on Raster. The new ray

17:34

tracing pipeline seems to be fast,

17:36

more performant, but it's still like

17:39

30% generation over generation uplift. And

17:41

without the DLSS for multi-frame generation

17:43

stuff, there's no kind of juice

17:45

here, which, you know, that feels

17:47

bad. Yeah. The other interesting thing

17:49

they talk about a lot is

17:51

that they switch the model. The

17:54

neural net model for upscaling and

17:56

deal SS4 from a convolutional neural

17:58

network, which is what they've been

18:00

using since the very beginning to

18:02

a transformer neural network Which is

18:04

what like chat GPT uses because

18:06

they've seemed like they had hit

18:08

the limits And they talked they

18:11

talked about this in that DF

18:13

interview with Alex. They also talked

18:15

about a lot in the white

18:17

paper They'd hit the limits of

18:19

what they could do with the

18:21

convolutional tech and we're we're finding

18:23

edges that they just couldn't train

18:25

around couldn't train beyond and cited

18:28

like things like chain link fences

18:30

as being less noisy and some

18:32

other like like difficult to render

18:34

places or places that are moving

18:36

quickly that we're providing ghosting other

18:38

weird artifacts the transform model supposedly

18:40

works better with. I'm curious to

18:43

see how what the difference actually

18:45

is in the real world like

18:47

there are extreme examples where DLS

18:49

2 and 3 doesn't look good

18:51

in certain education but in general

18:53

I think DLS looks pretty pretty

18:55

freaking solid. Like if it's available,

18:57

I'll always put it on quality

19:00

mode in a game that I'm

19:02

playing and I don't feel like

19:04

I lose anything. So I'm curious

19:06

to see like moment to moment

19:08

gameplay, how big of a difference

19:10

this shift in underlying AI networks

19:12

feels, models, not AI network. I

19:15

think the digital foundry guys were

19:17

talking about this on one of

19:19

the things that I think on

19:21

their their post like demo cycle

19:23

at CES. But like the like

19:25

the the frames when you're on

19:27

a high refresh rate monitor and

19:29

the frames are coming so fast

19:32

your ability to pick up the

19:34

little artifact that the artifacts are

19:36

getting more minor and they're coming

19:38

faster so you see them you're

19:40

less likely to to pick up

19:42

on them. So there's like it

19:44

it becomes a real like how

19:47

do we do image quality analysis

19:49

and benchmark testing on these cards

19:51

when so much of this is

19:53

outside of a traditional render pipeline

19:55

right like. I guess let's talk

19:57

about that later because that's a

19:59

big question. That's a really interesting

20:01

conversation to me. I have I

20:04

have some thoughts around that. So yeah, let's

20:06

go back to that. So other things

20:08

about the transformer model is it

20:10

runs on older hardware. It is

20:12

more expensive, but it's running a tensor

20:14

course. It is more expensive, but it's

20:17

running a tensor course. It's not like

20:19

it's taking stuff out of the raster

20:21

pipeline. It's just is is is the

20:24

part of your GPU that used to

20:26

be lightly used for DLS that's going

20:28

to be more used? I'm curious to

20:30

see is if they do roll that

20:32

out I think they said they're going

20:35

to do 40 series I'm not 100%

20:37

sure don't quote me on that and

20:39

they're looking into possibly do an

20:41

optimization for 30 series in the

20:43

future if that does come out I would

20:45

be really curious to see how moving

20:47

to the older tensor cores affects

20:50

memory usage in latency and GPU usage

20:52

overall because I think it would suck

20:54

into more of your stuff than you

20:56

would think so In the, in the

20:58

white paper, in one of the white

21:00

papers, they provided us, they broke down

21:02

generation by generation. And multi-frame generation,

21:05

the take one frame and generate four

21:07

off of it, or three additionals off

21:09

of it, is a 50 series only

21:11

thing. And that's because of the way

21:14

the tensor cores work on the new,

21:16

on the new GPUs in order to

21:18

have that happen in a timely manner.

21:21

Yeah. They had to do hardware stuff.

21:23

The single frame generation will continue to

21:25

be four series and five series. Ray

21:28

reconstruction goes all the way back to

21:30

2000 series to, is that ampere?

21:32

I can't remember. 2000 is ampere.

21:34

Yeah, and then, 30 was, 30

21:36

was ampere. Oh, it's a, pass, current,

21:38

Turing, Turing, Turing, Super

21:40

Resolution, beta, which is the

21:43

new transformer model, ups, upscaling,

21:45

goes all the way back

21:47

to 20 series. And DLAA goes

21:49

all the way, that's the anti-ali thing that

21:51

uses the DLS stuff, but is running NATO.

21:53

He'll be wrong on that then. Goes all the

21:55

way back to 20 series. And I don't know

21:57

if that's current or if that's coming, but that's.

22:00

like that's I'm real curious to see

22:02

how that stuff runs on a 20

22:04

series card because those those tensor cores

22:06

were not like they were not big

22:09

BP tensor cores. Oh no. Okay so

22:11

and it's it's an experimental thing the

22:13

tensor the the convolutional to transform a

22:16

model is experimental on most things in

22:18

order to enable you'll have to flag

22:20

it on the invidia app rather than

22:23

running it like there won't be a

22:25

drop down in the game to do

22:27

that. basically. Um, that's clunky, but it

22:29

makes sense. I mean, I think while

22:32

it's in, I think that's while it's

22:34

in beta is my understanding. Got you.

22:36

Um, the MFG stuff we talked about

22:39

a little bit. I kind of at

22:41

this point, I just want to see

22:43

what it looks like. Uh, digital foundry

22:46

who did a lot of spent a

22:48

lot of time with the demos at

22:50

C. At C. At C. S. Said

22:52

that it looks like they're their their

22:55

standards. before which were, hey, you want

22:57

this to run at 40 or 60

22:59

frames per second in order to turn

23:02

this on, remain true with the, with

23:04

the, so this isn't going to get

23:06

you from 15 frames a second to

23:09

60, it's going to get you to

23:11

from 60 to your monitor's refresh rate

23:13

at 240 or 360 or whatever. That

23:15

makes sense. Yeah, and I just want,

23:18

I don't want, I want to see

23:20

that, I want to see that running.

23:22

That's where I'm at with the MFG

23:25

because it sounds too good to be

23:27

true good to be true. One is

23:29

just a philosophical kind of way that

23:32

it seems to have approached at least

23:34

this generation. That's got me thinking about

23:36

graphics cards in general. Like I feel

23:38

like AI, up sampling, frame generation, who

23:41

else, who knows what else is going

23:43

to happen, neural rendering, that's new. Is

23:45

the future, and it's going to be

23:48

a big part of the future of

23:50

rendering graphics? Because honestly, like the GPUs

23:52

that we have already. are plenty powerful

23:55

enough to power 4K displays at gamer

23:57

speeds. Yes, you might have to. spend

23:59

a lot on a graphics card, but

24:02

the technology is there. So like the

24:04

issue with frame generation, especially multi-frame generation,

24:06

is the latency it's adds, which in

24:08

video reflex helps with. But it draws

24:11

off of your base frame rate anyway.

24:13

So in video has to keep adding

24:15

oomph. to make the base frame rates

24:18

go up in order to drive these

24:20

these super high frame rates with DLS4

24:22

MFG on and it's just got me

24:25

thinking like how much do they really

24:27

need to invest more in raster at

24:29

this point does it does it it

24:31

feels like it makes sense that they

24:34

invested so much in DLS4 in the

24:36

tensor cores because you only have to

24:38

drive so much additional raster performance use

24:41

that to unlock drastically higher raster performance

24:43

with MFG. I'm just getting philosophical about

24:45

graphics cards. I like it makes the

24:48

the conversation like this this feels like

24:50

again I haven't spent time with it

24:52

I'm curious like for me it really

24:54

all boils down to what it feels

24:57

like I don't care about the whole

24:59

semantic difference between fake frames and real

25:01

frames I know that people get really

25:04

upset about that but if I can't

25:06

like on the first-gen using a mouse

25:08

and keyboard I could perceive the difference

25:11

between frame gen on and frame gen

25:13

off in like a game like called

25:15

duty which supports it yes and runs

25:17

at 200 and something frames per second

25:20

on my computer right and the there

25:22

was there was a floatiness to the

25:24

input that you don't want on that

25:27

kind of game and and when I

25:29

did blind I didn't do double blind

25:31

so I don't have enough people to

25:34

do that but I did blind test

25:36

where had my daughter come in and

25:38

either toggle frame gen or not and

25:41

then I would play the game on

25:43

a specific level in a specific mode

25:45

and my kd would be would vary

25:47

by like a point five ratio when

25:50

i when i was testing it right

25:52

so you could feel it yeah it's

25:54

not scientific i couldn't feel that's the

25:57

thing is i had to i had

25:59

to like take my perceptions out because

26:01

Like my variance on KD is usually

26:04

like 0.1.2 on specific maps and specific

26:06

modes When I turn frame gen off

26:08

it went down turn frame gen on

26:10

it went down when I turned it

26:13

off it went up And and like

26:15

like like it's Yeah, it was weird.

26:17

It was weird. Some frames win games.

26:19

But at the same time, for something

26:21

like Indiana Jones, it's not a competitive

26:24

game, and I'm just playing to have

26:26

something that looks awesome and is fun.

26:28

And also that game does crazy, crazy

26:30

math, generating, showing the world and the

26:32

way the lighting works and all that

26:35

stuff is nuts. Like I would happily

26:37

have another, if I can, if I

26:39

can go from 60 frames a second

26:41

to 200 frames a second by flipping

26:44

on frame gen and at the

26:46

expense of a little bit

26:48

of snapiness on the aim,

26:50

I'm absolutely gonna do that

26:52

because the aim doesn't matter

26:54

in that game. Same thing for, you

26:56

know, God of war or horizon,

26:58

I'm just naming so first first.

27:00

Yeah, like, and then, and then, yeah,

27:03

like how, how, how, how, How do we test

27:05

AMD's frame gen versus invidious

27:07

frame gen versus Intel's frame

27:09

gen versus Qualcomm's frame gen

27:12

versus presumably some way Microsoft

27:14

will have frame gen? And yeah, anyway,

27:16

it's getting complicated. I go

27:18

back to I think it was last episode

27:20

somebody asked is average FPS still the

27:22

best way to measure it? And I've

27:24

been thinking about that that one

27:26

question, thanks to whoever asked it

27:29

hard ever since then because I

27:31

kind of align with hardware unboxed.

27:33

who are on the pulpit saying, they keep

27:35

counting these as frames, these aren't really

27:37

frames. They are, they improve the visual

27:39

quality, they're like motion smoother, it's

27:41

more of a visual quality thing

27:44

than a frames thing, which ties

27:46

into the feel and stuff like

27:48

that. So, how do you even measure

27:50

these things? Things are getting so complicated

27:52

now. Well, it's kind of like AA,

27:55

because if you think about AA, the

27:57

idea between annualizing is originally you would

27:59

render... four times the size and then

28:01

shrink it down and that smooths out

28:03

the jaggy edges right and what they're

28:05

doing with the with the frame generation

28:08

is kind of like that but for

28:10

time and stutter and for frame pacing

28:12

so they're basically what they're doing saying

28:14

hey we can make these three cheap

28:16

frames that we that we can deliver

28:18

evenly on a specific timeline and and

28:20

that's the other thing they change with

28:22

this generation is it's the GPUs controlling

28:24

the time the frame pacing for the

28:26

for the generated frames not the CPU

28:28

so that bottleneck is gone which makes

28:30

theoretically should make a big difference but

28:33

but yeah so they're saying hey we're

28:35

gonna get rid of shudder by making

28:37

these three cheap frames that go between

28:39

every expensive frame and we can dole

28:41

them out at a very predictable rate

28:43

so that you don't have that micro

28:45

shudder and if they actually can do

28:47

that then I'm really excited about this

28:49

because microstutter yeah like you can have

28:51

a game that runs at 250 frames

28:53

a 250 frames a second and if

28:56

it one frame out of 10 is

28:58

a 10x delay the game feels like

29:00

crap yeah yeah yeah so I don't

29:02

know we'll see this is like I

29:04

want to I want to sit down

29:06

in front of a nice monitor and

29:08

play a bunch of games and see

29:10

what it feels like that's that's the

29:12

only way to tell this stuff so

29:14

but here's a question I have since

29:16

we're kind of talking about the testing

29:19

aspect of it just uh just monitoring

29:21

chat like there is an array of

29:23

responses in in terms of like how

29:25

they feel about frame gen, how they

29:27

feel about latency, how they feel about

29:29

these other factors that go into how

29:31

good gameplay feels. And so I think

29:33

if we're talking about testing, it almost

29:35

sounds like it's yet another situation for

29:37

people who are benchmarking that it's just

29:39

going to be more complicated for them.

29:42

in order to present a fuller picture

29:44

to everybody. Because, I mean, we had

29:46

this when we started seeing more complexity

29:48

on the CPU side, right? So I

29:50

personally just, as someone who is not

29:52

testing GPUs and gets to sit back,

29:54

I always feel like it seems like

29:56

the traditional ways of testing are not

29:58

going to go out a window, but

30:00

you are going to have to add

30:02

like this extra layer to it so

30:04

that you get that data where you

30:07

can say like, okay, like we know

30:09

exactly how many quote unquote real frames

30:11

are being output by these cards that

30:13

if you're the kind of person who

30:15

really needs that, you can compare them

30:17

and at least it's like apples to

30:19

apples comparison. but then if you're someone

30:21

who's not sensitive to these things or

30:23

you are sensitive to these things then

30:25

here's the other like experiential plus the

30:27

numbers that go with that sort of

30:30

experience and it just sounds like I

30:32

don't envy anyone who's gonna have to

30:34

test this but it seems like to

30:36

satisfy everybody in chat who's looking at

30:38

reviews and trying to make a decision

30:40

for themselves plus the normies so to

30:42

speak who are gonna have to like

30:44

understand Well, this card gets 240 at

30:46

this, you know, resolution and settings and

30:48

this card was only getting 90. Like,

30:50

why is there this difference? And they

30:53

just go by numbers. Like, there's going

30:55

to have to be a way to

30:57

synthesize all that for them. And again,

30:59

I'll just say it again, because I

31:01

just keep thinking this over and over

31:03

again. I don't envy anybody who's going

31:05

to have to do this. But do

31:07

you see it differently though? Like, do

31:09

you see a way like a shortcut,

31:11

like a shortcut through a shortcut through

31:13

that? experience on your belt for this?

31:16

I think that's right. I think you

31:18

have to start with, I think you,

31:20

like anything else, you have to kind

31:22

of, you block and tackle the problem,

31:24

right? So you look at, it's all

31:26

built on a base of raster performance,

31:28

so you start with raster performance, and

31:30

then you test retracing performance, and then

31:32

you test. up sampling performance and then

31:34

you trust re-reconstruction performance and then you

31:36

trust frame generation performance and then you

31:38

compare frame generation on one piece of

31:41

hardware with frame generation on another vendors

31:43

piece of hardware and see who has

31:45

the best frame generation performance and pretty

31:47

soon we've done 350 videos and it's

31:49

only January so you know I'm only

31:51

laughing because of the pain because it's

31:53

actually It was it was bad when

31:55

it was only ray tracing and DLS1

31:57

now that there's so much stuff. There's

31:59

I think like Will was saying, it's

32:02

block and tackle, you got to

32:04

start with the foundation and then

32:06

you follow up with additional testing

32:08

from there. I think there will

32:10

still be some masochists who try to

32:13

offer comprehensive everything you want

32:15

in this, you know, review kind

32:17

of video or article, but that

32:19

will kill you. You don't get that long

32:21

to test these things and they take a

32:24

long time to test. So I think we're

32:26

going to start seeing even more of

32:28

like we see a lot of hardware

32:30

inboxed where they're like okay here's the

32:32

base review the next next day okay

32:34

we did these in ray tracing here's

32:36

the ray tracing results after that

32:38

you know we heard of this issue so

32:40

we're going to test this and then after

32:43

that okay now we did 50 games so

32:45

that you can see an even bigger thing

32:47

I think we're going to just start seeing

32:49

more stuff like that because it's just getting

32:52

complicated now. Like I love hardware and box.

32:54

This isn't a dig at them because I

32:56

think they provide an incredibly valuable service.

32:58

Yeah. But I think understanding what the

33:00

benchmarks are saying as as important as

33:03

having a whole buttload of benchmarks at

33:05

this point. And I agree. The DF

33:07

guy said this too. Like I think

33:09

that there I think that we have

33:11

to start talking about qualitative components here

33:13

because there is an experiential

33:15

aspect to this that for example.

33:17

when you're running a frame gen one

33:19

game your steam frame counter and frame

33:22

view are the tool that you use

33:24

to measure generated frames are going to

33:26

report different numbers now and one of

33:28

those is the is the real here's

33:31

what the raster is generating and one

33:33

of them is the raster plus frame

33:35

gen and that's only going to get

33:37

more complicated as frame gen becomes more

33:39

important so it's it's a real like Like

33:42

it's it's it's and it doesn't matter

33:44

if the if you have 240 frames

33:46

a second if the whole thing looks

33:48

like crap Then you might as well

33:50

run at 10 80 p and 60

33:52

hertz and have something that looks nice

33:54

as I look you know, what's Adams

33:56

smooth 40, you know, that's With the

33:59

kids crave. Yeah Yeah. I just really

34:01

hope Michael's off steps up and figures

34:03

out a way to standardize this someday,

34:05

like to make it easier for game

34:07

devs. If you could have more games

34:09

or all of these were available in

34:11

all three technologies available in a game,

34:13

like it'd be a lot easier to

34:15

test. But the Wild West nature of

34:17

these things makes it difficult. That would

34:20

simplify the matrix because then you're just

34:22

comparing against direct X upsampling versus. Sounds

34:24

like to me too though that like

34:26

I think you really touched on this

34:28

will that kind of like experiential reviews

34:30

like movie reviews TV show reviews reviewers

34:32

are going to have to get really

34:34

good about being clear about the things

34:36

they notice and don't notice experientially so

34:38

that even if you don't like this

34:40

is what I learned which is that

34:42

your job as a reviewer is to

34:45

be so clear about how your preferences

34:47

and experiences lie so that even if

34:49

somebody doesn't have your way of literally

34:51

viewing the world. They can use you

34:53

as a consistent like marker for how

34:55

they can predict their response will be.

34:57

Well, and I think we've had the

34:59

luxury of just being able to rely

35:01

on numbers and just presenting them and

35:03

just saying figure it out for yourself

35:05

and not necessarily have ourselves inserted into

35:08

the process. So this is, that sounds

35:10

different to me too. Well, and it's

35:12

a weird difference too, right? Because like

35:14

if you're talking about a film review,

35:16

it's about how did it make you

35:18

feel. There's physiological differences in all of

35:20

our abilities to detect changes in frame

35:22

rate and stutter and color saturation and

35:24

like like everybody's sensory sensorium is different

35:26

than everyone else's and we all live

35:28

on a spectrum where some people at

35:31

the old future days we had an

35:33

editor who could walk in be like

35:35

oh yeah that's under 60 frames a

35:37

second and he was right every time

35:39

right like he was just able to

35:41

see lower lower frame rates consistently. I

35:43

am not spring rate sensitive at all,

35:45

so I have to rely on what

35:47

the machine says. But I do get,

35:49

I'm so jealous of you. But I

35:51

do get, my, my first stoners make

35:54

me crazy. So like average frame rate,

35:56

as long as it's above about 50,

35:58

I'm fine. But if there's, if there's

36:00

frame sags in the, in the middle

36:02

of something that's running really fast, it

36:04

feels like it's running like 10 frames

36:06

a second to me. And, and it's

36:08

just in difference in how our brains

36:10

interpret visual input. And we have to

36:12

account for that and know, we have

36:14

to communicate that, and we have to

36:17

communicate that to communicate that to communicate

36:19

that to communicate that to communicate that

36:21

to communicate that to communicate that to

36:23

communicate that to communicate that to communicate

36:25

that to communicate that to communicate that

36:27

to communicate that to communicate that to

36:29

communicate that to communicate that to communicate

36:31

that to communicate that to communicate that

36:33

to communicate that to communicate that to

36:35

communicate that to communicate that to communicate

36:37

that to communicate that to communicate that

36:39

to communicate that to communicate that to

36:42

communicate that to communicate that to communicate

36:44

that. So and that's that's where frame-gen

36:46

is so good. I actually love frame-gen

36:48

I can't talk about multi-frame-gen yet obviously

36:50

but last original frame-gen like I am

36:52

someone who does feel latency like when

36:54

I'm moving my mouse and everything I

36:56

love how visually clean frame-gen makes things

36:58

look it really it helps with the

37:00

the pacing of it it's great me

37:02

I could feel the difference like Okay,

37:05

this says I'm going 200 frames a

37:07

second, but my mouse only feels like

37:09

it's going 100. So I could feel

37:11

that. And so I'm really curious to

37:13

see what my own hands on testing

37:15

of multi-frame generation finds. Yeah, like, if

37:17

this is, if this gives us a

37:19

world where everything feels like doom 2016,

37:21

then the frames are placed perfectly, I'm

37:23

there for that, right? Yeah. Let's see,

37:25

they talked about reflex two a tiny,

37:28

tiny bit. They didn't have even a

37:30

dedicated section to it, but they did

37:32

have demos of the finals. Apparently it's

37:34

also coming to valorant. They promised a

37:36

75% reduction in latency using reflex two,

37:38

using optical flow and motion vectors and

37:40

the same stuff, but better pipeline it

37:42

seems like on the newer hardware. It's

37:44

a little unclear to me where the

37:46

line is on what supports reflex two

37:48

and what is stuck on reflex one.

37:51

But the MFG, the multi-frame gen stuff

37:53

is not tied to reflex two. It's

37:55

working with reflex one apparently. Okay. And

37:57

then they also talked about ray reconstruction

37:59

a lot, which is interesting in that

38:01

it lets you get more. The basically,

38:03

it's using the tensor cores, the AI

38:05

hardware and the chip to get more

38:07

rays than the ray stuff can do.

38:09

So you get less, like if you

38:11

played, if you played ray tracing games,

38:14

you sometimes get a little bit of

38:16

visual noise that they often will smooth

38:18

out with other older techniques. Like it

38:20

almost looks like dots. And so the

38:22

ray reconstruction helps with that. And

38:24

maybe the noiser is so good. It's

38:26

crazy. Yeah, like. I think about what

38:28

control looked like on a 2080 and

38:30

I think about what a new game

38:33

looks like now and it's just like

38:35

the distance we've come in five years

38:37

is unbelievable. Yeah and back then control

38:39

was gorgeous. I loaded it up. Control

38:41

is gorgeous. I loaded up. I wanted to

38:43

see what it looked like last week when

38:46

I was getting ready to do this and

38:48

I was like yeah control still looks

38:50

really good. That's a good looking

38:52

game. All right. I don't want

38:54

to put to find a point

38:56

a point a point on it.

38:58

So this may be an indicator

39:00

of, sorry. Children, surrounded by children.

39:03

But I mean, okay, so I've

39:05

been doing this now. The

39:07

first video card I reviewed

39:09

for Maximum PC 100 years

39:12

ago was the G Force

39:14

DDR. So when they pulled

39:16

out that slide at the editors

39:19

day that's like, hey, here's

39:21

our first shader model with

39:23

the G Force 3. It's

39:25

like, oh, crap. Yeah, so, but like

39:27

in the, I guess what I'm thinking

39:29

about is in the ray tracing

39:32

era, we've gone from something that

39:34

was extremely rudimentary and could do

39:36

some nice looking reflections and like

39:39

a little bit of shading to

39:41

Indiana Jones running at with a

39:43

fully ray traced lighting model

39:45

in complicated indoor and outdoor

39:48

environments with multi light sources

39:50

with torches that you can

39:53

throw through the environment and

39:55

and like. Things that seemed like

39:57

magic five years ago like Hey,

40:00

the character's coat is shading the

40:02

area inside their, inside their jacket

40:04

when the light sources trigger is

40:07

just like, it's just table stakes

40:09

now, which is insane. Yes, it

40:11

is. It's insane. You've made me

40:13

wanna play Indiana Jones on my

40:16

PC. You should play it like,

40:18

I'm gonna go and tell you.

40:20

I know, like, Indiana Jones is

40:22

a little bit fraught as a

40:25

franchise for a whole lot of

40:27

reasons we don't need to get

40:29

into. This game, if you like

40:31

graphic stuff, is unbelievable and is

40:34

unlike anything I've seen out of.

40:36

The Star Wars game is interesting

40:38

and has a lot of similar

40:41

stuff going on, the massive Star

40:43

Wars game, but the game is

40:45

I think less good. The Indiana

40:47

Jones game is a real nice

40:50

combination of, yo, the graphics are

40:52

nuts and the game is really,

40:54

really good. So. Just to see

40:56

where everybody in a chat is

40:59

like sitting with like how they

41:01

feel about all this AI and

41:03

frame gen and the new stuff

41:06

like surrounding the traditional Raster performance

41:08

for Blackwell. What's the question? Is

41:10

it like, is it about, is

41:12

it, do we care about frames

41:15

being AI generated or Raster or?

41:17

Like where do we fall? Is

41:19

there a spectrum there? Is it

41:21

like our AI generated frames real

41:24

fake? Don't care as long as

41:26

it looks good. I don't know.

41:28

Yeah, something like that. What do

41:31

you think about all this DLS

41:33

and AI generative frames? I'm in,

41:35

I'm out. Don't care. Yeah, sounds

41:37

good. Alex, you got that? You

41:40

don't mind doing the... Okay. Okay.

41:42

We've got our top men on

41:44

it. Just while we're shifting gears

41:46

and gears there. Everyone go

41:48

watch gamers nexus 59 years found an edition tear-down video that they

41:51

did actually at CES with Invideo thermal engineer. It's freaking amazing. It's

41:53

great. It'll tell you everything you need to know about that

41:55

Like we were recommended it back

41:57

and forth reference it several times

41:59

in our own full nerd chat

42:01

at work. So go watch it

42:03

I learned stuff about vapor chambers.

42:05

I didn't understand before and like

42:07

they immediately answered my question about

42:09

how Oriented if orientation matters with

42:11

the wicking and all that stuff.

42:13

It's it's a super good super

42:15

good videos Steve Steve as always

42:18

knocked it out the park Can

42:21

we talk about the stuff from the

42:23

so they and editors days They also always

42:25

show some stuff that's not going to

42:27

ship in any games for a kind of

42:29

a long time Or will be in

42:31

limited games, but that is indicative of what's

42:33

coming in the future and a lot

42:36

of that was AI assists for things that

42:38

have not Not existed in the past,

42:40

right? Yeah, so like the neural shader thing

42:42

that they talked about which is basically So

42:45

what's the way to say this when you

42:47

when you build a material when you build

42:49

a game and you put a surface on

42:51

something You build in like unreal or unity

42:54

or whatever your game engine is you build

42:56

a material for it that material indicates like

42:58

everything from like the color and the and

43:00

the texture that what you think of as

43:02

a typical texture to like the The the

43:04

way the surface works like then the nobleness

43:06

of the surface. That's a normal map They

43:08

do how shiny it is which is specular

43:10

stuff they can convey even like all the

43:13

way down to like the sounds that it

43:15

makes when it's impacted by a bullet or

43:17

a foot or something like that and a

43:19

lot of those are shaders so that's so

43:21

like a lot of like the normal maps

43:23

and the The surface textures and things like

43:25

that are controlled by these little programs that

43:27

run on the GPUs called shaders and Typically,

43:31

it's fairly computationally expensive, right?

43:33

It's it takes a lot

43:35

of RAM because all these

43:37

little programs run in Jeep

43:39

in GPU memory so they

43:41

showed this neural shader thing

43:43

which is basically condenses a

43:45

big chunk of a material

43:47

of the materials that you

43:49

use into one calculation that

43:51

runs on the tensor cores

43:53

on the GPU and Also

43:55

increases image quality because you're

43:57

able to do more you're

43:59

implying more math from the

44:01

neural shader than you are

44:03

it's kind of it was

44:05

a little bit over my head. Complicated. Yeah,

44:07

but it solves both the, hey, we don't have memory

44:09

on GPUs and dedicated V RAM and it makes the

44:11

math, moves the math under a different part of the

44:13

GPU that typically shaders run, which frees you up to

44:15

do other things with the same units. And the thing

44:17

that I didn't get from their talk

44:19

but that I heard through scutel

44:21

butt is that Intel, AMD Qualcomm,

44:24

Qualcomm and Microsoft are all on

44:26

board to make this part of

44:29

a broader standard. and live outside

44:31

of like the RTX plug-in pipeline.

44:33

They have I think Vulcan extensions

44:36

coming for cross-platform games and for

44:38

stuff like Doom and you know

44:40

Ed Tech and all that. But

44:42

also the the the other three hardware

44:45

vendors being on board to

44:47

build this out is very

44:49

exciting to me. Yeah that was the

44:51

one question I had watching this

44:53

like I was watching it at

44:56

the keynote. I was in the

44:58

chat. I was like, this sounds super awesome.

45:00

The question is, will people actually use

45:02

it? Because in video has a rich

45:04

history of announcing super awesome new sounding

45:07

stuff like RTX storage or whatever, came

45:09

out with the 20 series. And then

45:11

just nobody ever uses it. And so

45:13

my concern is this sounds like it could

45:16

be awesome, but with AMD control

45:18

in the consoles and the handhelds,

45:20

like how many people are actually going

45:22

to use this? But then a couple

45:24

days ago Microsoft. put out a direct

45:26

X blog saying they're working on neural

45:28

rendering shading and is yeah it seems

45:31

like it's going to be an industry initiative

45:33

like you just said that that gives me

45:35

more hope for it. When I first saw

45:37

it I was deeply skeptical of developers actually

45:39

adopting it. So that's one that the

45:41

character artist and some of the technical

45:44

artists that I follow were like oh

45:46

this is really cool if this if

45:48

we get broad support for this this

45:50

will be amazing. This will let me

45:53

build something in the tools that I

45:55

know how to do and then essentially

45:57

It's not compression exactly, but like computationally.

46:00

Because the other thing that's important to

46:02

mention is the neural shaders, I think,

46:04

no, no, it's the textures that are

46:06

deterministic. Anyway, so yeah, the technical artists

46:08

were excited about that one. Like I

46:10

often buffer a lot of these announcements

46:12

through the real nerds that I know

46:14

that work in these spaces and whether

46:16

they're excited about it or not, because

46:18

they're the ones that are on the

46:20

front lines of whether this stuff is

46:23

gonna be used. Neural textures were another

46:25

one that was I thought was really

46:27

interesting and is actually. seems likely to

46:29

actually ship. This was just another, you

46:31

know, texture compression has been around for

46:33

a really long time. Consoles do it,

46:35

hardware has done it for years and

46:37

years. What they were promising with neural

46:39

textures is a seven to one further

46:41

compression over what you get with like

46:43

block compression using the AI algorithms to

46:46

expand and contract. That's deterministic, which means

46:48

it works the same every time. A

46:50

lot of AI stuff is non-deterministic, which

46:52

is not great for stuff like compression

46:54

because you want it to look the

46:56

same every time every time. And it

46:58

doesn't require decompression into memory. So like

47:00

typical texture compression, at least in the

47:02

old days, you'd store the textures compressed

47:04

on like, say, a DVD ROM that

47:06

was going into a PlayStation 2, and

47:09

then you decompress them into memory off

47:11

of that disk, and it would still

47:13

take the same amount of memory that

47:15

was an uncompressed texture in system ramp,

47:17

which is not gonna solve our, hey,

47:19

we don't have memory on video cards

47:21

problems. So this seems good, this seems

47:23

like something that will actually ship at

47:25

some point in the future. like have

47:27

some sort of neural compression in future

47:29

versions of direct X and maybe X

47:32

boxes. So that seems good. I like

47:34

that. I like, assuming they do get

47:36

picked up by the industry, I think

47:38

those are both good forward looking steps

47:40

for the industry that didn't get a

47:42

lot of attention. So I'm glad you

47:44

brought this up. The mega geometry one

47:46

is the last one. I mean, we

47:48

can talk about neural faces. I think

47:50

that one will, we can hit both

47:52

of them real quick. The mega geometry

47:55

one is like just. insane numbers of

47:57

triangles. Like just a lot a lot

47:59

of triangles. that they pitched as, hey,

48:01

this is the way we think rendering

48:03

is going to work in the future

48:05

for ray tracing and letting you do

48:07

ray tracing on for things that specifically

48:09

for things that move in the environment.

48:11

So typically, like in a ray traced

48:13

game, you don't have too many characters

48:15

on the screen at one time. You

48:17

don't have giant characters with billions of

48:20

triangles because the tessellation and the ray

48:22

tracing are kind of. The LODs are

48:24

kind of inherently incompatible with the way

48:26

rate tracing works because every time you

48:28

switch out of, LOD is when you

48:30

have a lower resolution of version of

48:32

something in the distance, it gets higher

48:34

resolution the closer you get to it.

48:36

And each time you kick it out,

48:38

you have to reset the pipeline and

48:40

you throw out a bunch of work

48:43

and that takes a whole bunch of

48:45

clock cycles and it's really inefficient and

48:47

it causes things like microstutter. This is

48:49

a way to keep all of the

48:51

geometry in. memory loaded at one time

48:53

without having to tessellate or without having

48:55

to LOD out so that you don't

48:57

have to throw out those ray tracing

48:59

calculations as you get closer and further

49:01

away and you need higher and lower

49:03

levels of detail, which seems really good.

49:06

It does. I found it interesting that

49:08

it tied into I believe Alan Wake

49:10

2 was the only game they talked

49:12

about for that section. Yeah. So I'm

49:14

curious to see Alan Wake 2 is

49:16

pretty much the only game I could

49:18

think of that uses an immediate feature

49:20

called. which lets you cram tons of

49:22

stuff on the screen at once. Basically

49:24

the non-ray traced version of what you

49:26

said. So I'm just curious to see

49:29

if this feature gets adoption pick up

49:31

or if it requires using those mess

49:33

shaders and is that something we're gonna

49:35

see more of in the future. And

49:37

the way a lot of this stuff

49:39

that rolls out is often it'll roll

49:41

out with a partner, an invidia partner

49:43

title first like Alan Wake or Indiana

49:45

Jones or whatever. as an RTX feature

49:47

for in-video user, for in-video graphics card

49:49

owners. And then everybody gets together and

49:52

they all talk and everybody talks to

49:54

Microsoft and then you see some version

49:56

of this that comes in as an

49:58

API for the next version of direct.

50:00

and then the other vendors tie into

50:02

that and eventually it's in the Xbox

50:04

is the kind of circle of life

50:06

for this stuff. I hope it's true for

50:08

this one because again I think this

50:11

sounds awesome. I think the mega geometry

50:13

patch for Allen Wake is supposed to

50:15

come out shortly either with release or

50:18

shortly after I don't know. And then

50:20

the neural faces is the one. So

50:22

in these announcements, there's usually like a

50:24

few that seem like they're really good

50:26

and they're going to ship and we'll

50:28

see them in games in like three

50:30

years. And then there's one that just

50:32

seems like complete hooey. And the neural faces

50:34

one was that one for me. Yeah, it doesn't

50:36

feel the same. Yeah, it's basically they

50:39

were, it looked like they were generating

50:41

player faces that were generated by character

50:43

artists and using them as a seed.

50:45

generating new faces with a generation a

50:47

generative AI model and then deep-faking that

50:50

face back onto the model in the

50:52

game so you had a face that

50:54

looked like a real face but it

50:56

was not it was it looked bad

50:58

the the demo that they showed at

51:00

the keynote look bad I haven't seen

51:03

it obviously running in real life

51:05

yeah I'd unen impressed question I

51:07

feel like the invidious teachers hmm

51:09

what's in that box right there we'll

51:11

have to say we'll go get there okay

51:14

I didn't know if we were getting there

51:16

or not. Yeah, no, I'm not going to

51:18

forget about the box. The box is

51:20

here. Just check. It's just sitting

51:22

here. For audio listeners, we have

51:24

a brown cardboard box just sitting

51:27

quietly next to Will, and we

51:29

have gotten at least one nudge

51:31

in chat from one Adam Patrick

51:33

Murray asking what's in that box,

51:36

Will? Look, it's not Gwyneth Paltrow's

51:38

head. It's boiler, I guess. To

51:40

me, the part where in video

51:42

features haven't been as exciting to

51:44

me, have been when they've leaned

51:46

heavily on generative AI, like that.

51:48

Like I love all of this other

51:50

stuff that we've talked about today,

51:52

but for me, the neural faces

51:54

or whatever they called it, also

51:56

fell flat. For me, like, I like the

51:59

flexibility. of, I forget the name,

52:01

the AI characters where they can talk

52:03

to you with generative AI and game,

52:05

Ace, Ace, Ace. I've actually played with

52:07

that. It worked better than I thought

52:10

it would when I played with it

52:12

last CES, but it still feels like

52:14

there's just a slight delay as it

52:16

thinks itself out and starts answering you

52:18

and stuff. And to me, I can

52:21

understand why invidious pushing so hard on

52:23

these things, but when they're turning to

52:25

generative AI rather than these other behind

52:27

the scenes AI uses, I think is

52:29

where like like like. some of the

52:31

magic starts to fall away from me.

52:34

That's exactly it for me is like

52:36

generative AI to make texture sizes seven

52:38

times smaller. That sounds fantastic. Generative AI

52:40

to make to upscale a 1080P render

52:42

to 4k and make it look as

52:44

good as a 4k render would at

52:47

a tenth of the frame rate or

52:49

10 times the frame rate. That sounds

52:51

great. I love that. Is that all

52:53

day? Yeah. Generative AI to have like

52:55

the problem every time I've done that

52:58

ace demo. It just becomes me wanting

53:00

to break the AI, which is not

53:02

a fun game. I mean, it's a

53:04

fun game for a little bit, but

53:06

then it's not fun pretty quickly after.

53:08

So anyway. You want to talk about

53:11

unboxing? You want to unbox? We have

53:13

a box here. We can unbox the

53:15

box. What's in the box? What's in

53:17

the box? Unbox the box. I wanted

53:19

to shout out in video because they

53:22

did a really nice job with packaging.

53:24

This is a 4090 box. It's mighty

53:26

and big. Wow, I can't believe you

53:28

actually had it just ready to go

53:30

for show and tell just to emphasize

53:32

this point. Yeah, it's big, it's huge,

53:35

there's a bunch of dead space which

53:37

I guess is good for shock. shock

53:39

protecting the board. And in fairness, I

53:41

thought that this was made with plastic

53:43

foam around the inside edges. It's actually

53:45

like corrugated cardboard around the edges patting

53:48

it out. So it's probably you can

53:50

just chuck it in the recycling bin,

53:52

which is which is important these days.

53:54

No, there's foam in there when I

53:56

at least some of them when I

53:59

had to empty out my house. after

54:01

it burned down last time. I had

54:03

a bunch of these boxes in my

54:05

basement that got soaked, so I had

54:07

to throw them away. And I could

54:09

only fit two of those boxes per

54:12

trash bag because they were tons of

54:14

foam and they're huge. Ooh. Okay, so

54:16

this one's cardboard. You open it up.

54:18

Here I mean, it doesn't angle automatically

54:20

like the 49. Which was the real

54:22

defining feature. But you lift it up

54:25

and then these sides kind of slide

54:27

out. And you lift

54:29

out the dog bone, which is contained

54:31

contains a 50-90 inside, oh, it's really

54:34

heavy. It looks like a milkbone, it

54:36

does. It's like a 12-pound box. And

54:38

then it's held together by just little

54:40

flips, you flips, which is pretty good.

54:42

Hold on, this one stuck mine. The

54:45

one I had didn't stick like this.

54:47

Is that just like layers of cardboard?

54:49

It is, it looks like layers of

54:51

cardboard to me. A corgated cardboard, yeah.

54:54

Interesting. That's hugely better for the environment.

54:56

I like, I like that change a

54:58

lot. Yeah, I thought the 40 series

55:00

were real flashy, but wasteful. Yeah, it

55:03

felt the 40 series, I mean, you

55:05

were buying a ridiculously expensive video card,

55:07

then you felt like you were also

55:09

kicking the environment as you were unboxing

55:12

it, so. I mean, you're already still

55:14

kicking the environment by using the card,

55:16

but you know. And then the lid

55:18

just kind of comes off, and there's

55:21

a 1590 inside. We move the. The

55:24

other box it's a it's a

55:26

they did a nice job on

55:28

packaging So we can talk about

55:30

on boxing I was surprised My

55:32

12 volt high power was a

55:34

450 watt one so I had

55:36

to go find a 600 watt

55:38

one to plug in I didn't

55:40

realize that they were different 12

55:42

volt high powers I didn't either

55:44

I want to shout out They

55:46

have a good the the the

55:48

octopus adapter for the 12 volt

55:50

high power to the four PCI

55:52

express powers. It's, you probably not

55:54

going to see it on the

55:56

video here, but it's got. like

55:58

a woven braided braided thing. So

56:00

it's nice and floppy. And then

56:02

where it comes out of the

56:04

top is in a good spot

56:06

to not put stress on the

56:08

on the connectors even with the

56:10

way of the four PCIE connectors.

56:12

I was very excited to see

56:14

them move to the braided cables.

56:17

I actually don't like those adapters.

56:19

They have to ship them. I

56:21

understand why they ship them. Most

56:23

people don't have a 12 volt

56:25

power power supply. But the last

56:27

two generations. Not only were the

56:29

cord short, but they were like

56:31

inflexible. So they looked ugly as

56:33

heck if you had a clear

56:35

panel on your case, which 95%

56:37

of people who listen to full

56:39

nerd probably do. And I hated

56:41

them. They were con in every

56:43

single one of my reviews. That

56:45

one not only is braided so

56:47

you could bend it more. It

56:49

looks a little bit longer too.

56:51

So if that's the case, I'm

56:53

excited about it. Yeah, they're pretty

56:55

long. They're like six inches, it

56:57

looks like. And also the connector

56:59

is trunkier. Like it's it's not

57:01

it's got more stress relief coming

57:03

out of the connector which I

57:05

which I appreciate. I think I

57:07

think with this one like you

57:09

might have had an 850 watt

57:11

power supply before that wouldn't have

57:13

12 12 volt power I don't

57:15

I think this is the minimum

57:17

is a thousand watts now and

57:19

I bet that people are going

57:21

to be doing some power supply

57:23

upgrades if they're going to a

57:25

50-90. If you're buying a $2,000

57:27

graphics card yeah. Pony. The biggest

57:29

problem I had installing this thing

57:31

the real problem is that we

57:33

can't talk about plugging stuff in

57:35

any more performance or anything like

57:37

that I'm just gonna say going

57:40

from a three slot card to

57:42

a two slot card I had

57:44

to go find the little thing

57:46

the adapt the slot cover because

57:48

I haven't had a slot cover

57:50

in my third GPU slot in

57:52

like probably since a 2080. Yeah,

57:54

it's been forever. Yeah, so I

57:56

was like, well, where's the slot

57:58

cover, man? I was like. Okay

58:00

I had to reverse look I went

58:02

on a journey I dug I went I

58:04

went through three different motherboard boxes

58:06

I had to go back in

58:08

time I had to like I

58:11

had to Colombo it I was

58:13

sitting there with a lollipop and

58:15

I was like okay what would

58:17

I have been doing in 2020 when

58:19

I got a 3080 TI and and

58:21

I went back in time and I

58:23

was like boom here it is I

58:25

got a I got a slot cover so

58:27

I found one. Very financially privileged thing

58:29

or I'm like, you know what they they

58:32

sell generic slot covers I'm just gonna buy

58:34

a set and I'm just gonna put that

58:36

on there I won't see that it doesn't

58:38

match. I want to have the same cuts in

58:41

the in the in the breathing holes Sometimes

58:43

sometimes Colombo and it just

58:45

takes a lot of brain energy like it's

58:47

got to be right. You can't you can't

58:49

half-ass. Ooh Adam ouch You started this, this

58:51

is on you, I'm calling HR. No, no.

58:53

This is a hostile work environment. No,

58:55

it was Adam said in chat, he

58:57

knows, wow, Columbo, he's old. I'm a

59:00

big fan of the Roflowe Columbo

59:02

movement to get Mark Ruflowe to

59:04

reboot Columbo, because he would be

59:06

perfect for that. He would be a great, he

59:08

would be a great choice for that. Yeah, he's

59:10

got the right, the right vibes and right

59:13

energy for that. So how so do we want

59:15

to talk about how we're like how we

59:17

hit the testing stuff? Do you think we

59:19

got enough? Do you have more thoughts on

59:21

that? Because I had a note at the

59:23

end to come back to testing and how

59:25

how you I want to talk about it

59:27

earlier. Okay, I talked about it. We do

59:29

have we don't have we didn't close the

59:31

poll yet, but after just 16 minutes, we

59:33

already have a hundred and fifty votes. That's

59:36

a lot of votes. Do we want to even

59:38

just look at it now? Yeah, okay. So I

59:40

don't know if this will be

59:42

exactly final, but the question is

59:44

are AI frames something you want

59:46

to use in your games? And

59:49

we have yes at 35-ish percent.

59:51

I just saw it go up

59:53

36 and go back down literally

59:55

in the second that took me

59:57

to read this. No, at 49

59:59

percent. and heck no at 16% I

1:00:01

wow wow it's not what I was

1:00:04

expecting people people hate the I was

1:00:06

expecting a hey yeah if they look

1:00:08

fine I'll totally turn them on but

1:00:10

I I actually wonder so you know

1:00:13

our audience is obviously full of people

1:00:15

who are very passionate about Harvard performance

1:00:17

and like really like to know you

1:00:19

know how the you know how it

1:00:21

goes when you put the pedal to

1:00:24

the metal kind of thing, right? But

1:00:26

I would be super curious if we

1:00:28

were to pull people who just kind

1:00:30

of buy a graphics card the way

1:00:33

that they just like buy a toaster

1:00:35

because they just wanted to do the

1:00:37

thing, like what their responses would look

1:00:39

like. Because I have a feeling that

1:00:41

it would be a different breakdown, but

1:00:44

I don't know without the actual data.

1:00:46

I like we should just go to

1:00:48

Best Buy on 5090 Log Day. You're

1:00:50

just standing there with clipboards. Would you

1:00:53

like to do a survey? Yeah. No,

1:00:55

no. I just want to ask you

1:00:57

about fake. We're not selling anything. No

1:00:59

political. Mostly. Except on the credit. Yeah.

1:01:01

Many people are missing out. I love

1:01:04

fake frames. Feel frame Jen is good.

1:01:06

In some cases it's not. Give it

1:01:08

to try. Turn off your frame counters.

1:01:10

Give it to try. See how it

1:01:13

feels. See how it feels. See how

1:01:15

it feels. See how it feels. Just

1:01:17

my plea. Just my plea. Just my

1:01:19

plea. Just my plea. Just my plea.

1:01:21

in multiplayer, like single player, I'm 100%

1:01:24

fake frames, multiplayer, a little less so,

1:01:26

I'm gonna say. I never touch it

1:01:28

in multiplayer, I would never touch it

1:01:30

in multiplayer. I do touch reflex every

1:01:32

time I can. Reflex is actually my

1:01:35

favorite in video technology. Yeah, and video,

1:01:37

reflex is the one thing I go

1:01:39

in, it's become the thing that I

1:01:41

go in and turn on in pretty

1:01:44

much every game. Yeah. I often, this,

1:01:46

the fake frames thing seems like, hey,

1:01:48

I haven't tried this yet. I haven't

1:01:50

been exposed to it in its reality

1:01:52

situation a lot of the time. So

1:01:55

anyway, wow, we just got 21 more

1:01:57

votes. People are like in the last

1:01:59

three minutes, you had 21 more votes.

1:02:01

It actually dropped to 34% for yes.

1:02:04

Oh, it's now 50% for no and

1:02:06

16% for heck no. So we have

1:02:08

a lot of skeptics among us. Voting

1:02:10

matters. Right now. Yeah. So I guess

1:02:12

that's, let's put a bow on the,

1:02:15

on the in-video editors day stuff. It

1:02:17

was, there's a lot more stuff there.

1:02:19

They talked about a ton. There was

1:02:21

a bunch of AI stuff that is.

1:02:24

I don't know. I watched the presentations

1:02:26

and read the white papers and I

1:02:28

was like, yeah, I don't know about

1:02:30

some of this stuff. A couple things

1:02:32

seemed interesting. We just spoke about all

1:02:35

this stuff. I agree it's about time

1:02:37

to move on, but if you, because

1:02:39

we were, like we are the full

1:02:41

nerd, if you want to learn more

1:02:43

about this stuff, like invidia does a

1:02:46

great job at putting those white papers

1:02:48

up on this website. Like I think

1:02:50

a lot of them are already active

1:02:52

on the website before these launch. So

1:02:55

if you want to go dig deep

1:02:57

into the stuff, go read the white

1:02:59

papers and then you might learn something.

1:03:01

There's a lot of good interviews. The

1:03:03

interview, the digital foundry interview with the

1:03:06

AI, with their AI expert, is really,

1:03:08

I learned a ton. Steve's breakdown of

1:03:10

the cooling of the cooler on the

1:03:12

50-90 is fantastic. Highly recommended all around.

1:03:15

Should we talk about B570 reviews? Intels?

1:03:17

We should. I may not be able

1:03:19

to stay for the whole of it

1:03:21

because unfortunately I have to pop off

1:03:23

light around 1230. Okay. But we should

1:03:26

go. So what's the story on B570

1:03:28

reviews? Pretty good? Or we didn't, we

1:03:30

don't, I think we missed the embargo

1:03:32

on this just because we had competing

1:03:35

things coming in and only limited time.

1:03:37

We had things too. But you're supposed

1:03:39

to start the year off light? It's

1:03:41

been a very hectic start to the

1:03:43

sheet. Yes. But yes, sorry until we

1:03:46

did miss it. Arts Technica, I read

1:03:48

the Arts Technical Review and they, they're,

1:03:50

their deck kind of gives the whole,

1:03:52

or the headline gives the whole thing

1:03:55

away. They said, hey, $ 219, it's

1:03:57

the cheapest good graphics card competitive with

1:03:59

a 4060. with 10 gigs a RAM.

1:04:01

That seems pretty good. Yeah,

1:04:03

yeah, that's pretty much, it's

1:04:05

exactly what I expected

1:04:07

it to be. Like, you know, it's like

1:04:10

80% of the B580, or you know,

1:04:12

something like that. It chops off a

1:04:14

block, so it has, you know, fewer

1:04:16

cores, it has two gigabytes, fewer

1:04:19

gigs, so 10 gigs. And it

1:04:21

performs about what you expect.

1:04:23

If you're in a super tight

1:04:25

budget, like to me, that's

1:04:27

a more compelling option than.

1:04:30

you know some of the ones that have

1:04:32

been around for a while like a 7600

1:04:35

or a 6600 also was a big budget

1:04:37

card for a while I'd rather get this

1:04:39

than that but really if you can

1:04:41

save an extra week and save up for

1:04:43

B 580 that's the way I would go

1:04:45

but if you can't get a B 580

1:04:47

because they're sold out everywhere this looks like

1:04:50

a good budget option I just would try

1:04:52

not to pay more than $220 $230 max

1:04:54

What's the, did you get any

1:04:56

more insight into what's going on with

1:04:59

the, it seemed like there were performance

1:05:01

problems with machines that maybe didn't have

1:05:03

resized bar support or people hadn't turned

1:05:06

that on? That has always been the case

1:05:08

even with the first gen art cards. It

1:05:10

still is now, I can say less egregious

1:05:12

now, pretty much any computer sold this

1:05:14

decade has resizable bar and even

1:05:16

back even further were former updates.

1:05:18

Is it issue with the first

1:05:21

gen cards not with this one?

1:05:23

The bigger issue going around now

1:05:25

is that apparently Intel's driver

1:05:27

has a lot of overhead. So if

1:05:29

you have a slower CPU, an

1:05:31

older slower CPU, so if

1:05:33

you have like an old rise

1:05:35

in 2,600 or 3,600 system,

1:05:38

and you play CPU heavy

1:05:40

games in particular. But

1:05:42

factorial, some open world

1:05:44

games at CPU heavier

1:05:46

you would think to. They can

1:05:48

have. much less performance than you

1:05:51

would see with a modern faster

1:05:53

chip. So if you're building a

1:05:55

new system, I would think you'd

1:05:57

be fine unless you're up in

1:05:59

493. But if you're looking to

1:06:01

upgrade an older system, even though

1:06:03

the B570 and B580 look great

1:06:05

in the vacuum, if you have

1:06:08

one of those older Rison5 chips,

1:06:10

Intel Corei 5 chips from the

1:06:12

same era, you probably want to

1:06:14

look to Radiana and Vidia instead.

1:06:16

Okay. That sounds good. That's what

1:06:18

you want to hear. It's a

1:06:20

little weird to me that there's

1:06:22

a pricing that the price on

1:06:24

this is so close to the

1:06:26

founders edition or the Intel edition

1:06:28

B580. Because it's like a $30

1:06:31

difference. But also those cards, I

1:06:33

guess, are going to be hard

1:06:35

to get, come by. So. They're

1:06:37

all sold out right now. The

1:06:39

ones that you can buy are

1:06:41

from Gunnier, Gunner. And they're shipping

1:06:43

from China and they're like 350

1:06:45

bucks at which point don't buy

1:06:47

those. So B580 is out of

1:06:49

stock. You could still, I found

1:06:51

one B 570 card. But yeah.

1:06:54

This Ray and D has done

1:06:56

this for a while as a

1:06:58

way to get rid of bin

1:07:00

dies. Like they've often done like

1:07:02

the 7700xT is came out shortly

1:07:04

after the 7800xT for like $50

1:07:06

less for example and it's just

1:07:08

a way to kind of upsell

1:07:10

but also get rid of the

1:07:12

old dies. Okay. Defected dies. Yeah.

1:07:15

Yeah. Like better than throwing them

1:07:17

out. Yeah. But it's a good

1:07:19

card. B580, 30 bucks more if

1:07:21

you can find one, probably a

1:07:23

better investment, but it's great to

1:07:25

see a budget option like this

1:07:27

for gamers. Ten 80P, right, is

1:07:29

where your sweet spot is here

1:07:31

probably. I read a bunch of

1:07:33

reviews and then I read a

1:07:35

meta review that contextualizes them all.

1:07:38

And just like the B580, the

1:07:40

B570 gets better at higher resolutions,

1:07:42

for whatever reason it's architecture and

1:07:44

its memory. configuration just is better

1:07:46

when you have rate tracing and

1:07:48

1440 P. But if you're going

1:07:50

to be playing at 1440 P

1:07:52

I would I would definitely try

1:07:54

to save up for a B

1:07:56

580. That extra performance is worth

1:07:58

it. sense. Do you want to

1:08:01

take a couple of cues and

1:08:03

turn them into A's before we

1:08:05

have to sign off? Elena, do

1:08:07

you have anything to add on

1:08:09

the B570? I'm sorry. No, I

1:08:11

mean, it seems pretty straightforward, which,

1:08:13

I mean, in a good way,

1:08:15

which is nice, right? We were

1:08:17

just saying at the big game

1:08:19

of the show that we're like,

1:08:22

we like things that seem uncomplicated.

1:08:24

So I guess if you have

1:08:26

a question. for us here at

1:08:28

the full nerd. You can post

1:08:30

it in the YouTube channel or

1:08:32

you can go into the discord

1:08:34

where the link is below in

1:08:36

the description for our YouTube channel

1:08:38

and there's a full nerd questions

1:08:40

channel in there that you can

1:08:42

post questions in and we will

1:08:45

turn them into answers for you.

1:08:47

Our first question is from Ivan

1:08:49

R over in Discord and he

1:08:51

says, hey this is for Brad,

1:08:53

can MFG multi-frame gen DLSS4 make

1:08:55

G-force now better? That

1:08:59

is a really interesting question. That is

1:09:02

a really interesting question. I think it

1:09:04

would depend on one if they upgrade

1:09:06

the server hardware to 50 series. Currently,

1:09:08

actually, we just wrote about it today.

1:09:11

G force now is unavailable in a

1:09:13

lot of places. Like you can't open

1:09:15

a free account. You can't open a

1:09:17

six month account unless you want ultimate

1:09:20

addition. So they might be updating their

1:09:22

infrastructure right now. I forget this DLS

1:09:24

work on. DLS3, does that work on

1:09:26

the ultimate edition? If DLS3 works on

1:09:29

G4S now, I have no doubt that

1:09:31

multi-frame generation would work in the same

1:09:33

way. I'm just not sure off the

1:09:35

top of my head if that supports

1:09:38

DLS3. Yeah, that, well, and the way

1:09:40

the frame gen stuff works, you have

1:09:42

to have, like, you have to be

1:09:45

able to talk to the monitor. So

1:09:47

I don't know if the G4S now

1:09:49

stuff talks to the monitor. It does.

1:09:51

Okay. Yeah, they do in some way.

1:09:54

I forget the hairy details, but they

1:09:56

do. They adapt to the monitor. Okay,

1:09:58

that's interesting. Well, there you go. But

1:10:00

I'd assume it'll take a while to

1:10:03

roll out because they're going to move

1:10:05

those. I think my understanding is the

1:10:07

way the G4 Now stuff works is

1:10:09

you get, they get, it's basically it

1:10:12

runs in the slack on their cloud

1:10:14

compute platform. So when they're not selling

1:10:16

cloud compute to somebody who wants to

1:10:18

do massive AI training or whatever, then

1:10:21

you're playing your games on the spare

1:10:23

cycles that are left over. So it

1:10:25

may not be that the G4s now

1:10:27

load is so high. It may also

1:10:30

be that the the the compute load

1:10:32

is so high that they don't have

1:10:34

extra slack right now, right? Yep, that

1:10:37

would make sense. I just looked and

1:10:39

G4s now does support DLS3. So I

1:10:41

would expect assuming they switch over the

1:10:43

hardware that DLS4 would also be supported.

1:10:46

Here's one from fly. Oh, go ahead.

1:10:48

Oh, is the next one about frame

1:10:50

generation because I saw a good one

1:10:52

in chat. Okay, so this is from

1:10:55

Jesse Vivano. How would frame generation help

1:10:57

at all in fighting games? Real frames

1:10:59

have to be delayed to fit the

1:11:01

AI frame somewhere. I don't think you're

1:11:04

gonna see a frame generation in fighting

1:11:06

games because of the way frame pacing

1:11:08

works in fighting games. Oh, the frame,

1:11:10

the frame basin is so important. Nobody

1:11:13

would ever use that competitively. Yeah. Well,

1:11:15

and I mean, it's because they're usually

1:11:17

locked at 60 frames per second so

1:11:20

that you so that people who see

1:11:22

the frames can time their moves and

1:11:24

stuff like that. So yeah, it makes

1:11:26

such a difference when you're doing combos

1:11:29

and stuff like that infamous Justin Wong,

1:11:31

I believe his name was the the

1:11:33

Chen Lee where he blocked the Chen

1:11:35

Lee's kicks that moment. Like there's no

1:11:38

way you could do that if you

1:11:40

had three AI frames between every one

1:11:42

of those frames. Yeah, I mean, but.

1:11:44

I mean, look, we all know that

1:11:47

people are really fighting games are just

1:11:49

button-mashing. So, you know, it's, uh... Wow,

1:11:51

wow, we get spicy in this, the

1:11:53

Q&A section. I'm kidding. I used to,

1:11:56

I used to mop up betting on

1:11:58

our... games back in the day at

1:12:00

the mall. That's how old I am.

1:12:02

Look, I was never allowed to go into

1:12:05

the mall where the into the arcade where

1:12:07

the betting happened. That was at the bowling

1:12:09

alley. The one in the mall was the

1:12:11

was the was the family one. The name

1:12:13

of the gold mine because they knew what

1:12:16

was up. Yeah, it's interesting because

1:12:18

that's the one place that the AI

1:12:20

frame gen stuff isn't going to land

1:12:22

because those like those games were just

1:12:24

built to run it at a locked

1:12:26

lock frame rate. Yeah, like it totally

1:12:28

makes sense. If you wanted to do

1:12:30

it and didn't care competitively turning it

1:12:32

on in a shooter, I guess you

1:12:34

could if you were just maybe not

1:12:36

a competitive shooter. But yeah, it

1:12:39

makes no sense for fighting games. The

1:12:41

competitive scene is what makes or

1:12:43

breaks fighting games. I just don't see

1:12:45

them adding that. Creativity

1:12:48

404 asks, Invidia advertises up to

1:12:50

three media encoders on the 50

1:12:53

series. On the 40 series, only

1:12:55

4070 TI and up had double

1:12:57

encoders. Did invidia give any more

1:13:00

info on which tier will be

1:13:02

enjoying this triple encoder? I

1:13:05

don't actually think I saw a breakdown

1:13:07

on which one gets what. I think

1:13:09

I read somewhere. It's expected just

1:13:11

on the 50-90, but don't

1:13:14

quote me. It would make sense for them

1:13:16

to put that in the 50-90. I was

1:13:18

kind of curious. What do you think people

1:13:20

are going to use triple encoders

1:13:22

for Brad? I don't know. Like, I'm sure

1:13:24

there's some complex thing, but like you

1:13:27

were saying last time we talked, like

1:13:29

when you're doing the multi-stream to Twitch,

1:13:31

it might help out? Yeah, we did

1:13:33

talk about this last week. This is

1:13:36

a problem with doing too many podcasts.

1:13:38

Yeah, I can never remember which when

1:13:40

I had the conversation I had the

1:13:42

conversation with. But yeah, I think

1:13:45

that would be the main driver

1:13:47

Who asked the question if if

1:13:49

it's a live question shout out

1:13:51

while you're asking to the

1:13:53

comments and we'll have a

1:13:55

reason creativity 404 asked that

1:13:58

one Let's see flying shoot has

1:14:00

a fun one here. When eating

1:14:02

string cheese, are you the type

1:14:04

of person that peels it or

1:14:07

bites straight into it? And then

1:14:09

they have some other some other

1:14:11

context that I'll add after we

1:14:13

gives their answer. I'm a peeler.

1:14:16

Straight peel. Just I open it

1:14:18

up and I just strings. It's

1:14:20

called string cheese for a reason

1:14:22

y'all. I respect that. Brad, there's

1:14:25

lactate pills. Yeah, I try not

1:14:27

to. I just avoid them for

1:14:29

the most part. Wow. What about

1:14:31

you, Elena? Okay, so primarily I

1:14:34

will peel it, but there are

1:14:36

some times where I'm so desperately

1:14:38

hungry that I do not have

1:14:40

the patience to painstakingly peel it,

1:14:43

even though it is without a

1:14:45

doubt, the much more texturally preferable

1:14:47

way to eat string cheese, because

1:14:49

if you bite into it, like

1:14:52

it's just a hunk of normal

1:14:54

cheese, it's gross. like straight up

1:14:56

it's gross texturally it is not

1:14:58

good but sometimes needs the needs

1:15:01

of the stomach are greater than

1:15:03

the wants of the mouth Alex

1:15:05

string cheese peel or bite string

1:15:08

cheese is disgusting you should leave

1:15:10

it on the shelf Flying shoes

1:15:12

follow up to that as my

1:15:14

feeling is if someone bites into

1:15:17

the string cheese, they're more likely

1:15:19

to be a psychopath. So there

1:15:21

you go. I agree. Like psychopath

1:15:23

tendencies. Yep. Yep. However, Adam went

1:15:26

full out saying that as a

1:15:28

kid, he would peel it and

1:15:30

as an adult, he just bites

1:15:32

it. So let's take away what

1:15:35

you want from that's. I mean,

1:15:37

Adam, Adam does eat gum off

1:15:39

of the floor trade shows. So

1:15:41

take that for information with the

1:15:44

context. Let's see, I'm looking for

1:15:46

more questions. Jesse Viviano said, have

1:15:48

you heard that in video publishers...

1:15:50

vulnerability report on drivers. I saw

1:15:53

new stories about that, but I

1:15:55

just used that as an excuse

1:15:57

to update my drivers and didn't

1:15:59

actually read what the vulnerabilities were.

1:16:02

Yeah, same here. I saw the

1:16:04

new stories and I didn't have

1:16:06

time to look into it. It's

1:16:08

been a very busy month. It

1:16:11

has been a very busy month.

1:16:13

I think that's, somebody else saw

1:16:15

something else. Let's see, Devina said

1:16:17

earlier, Devina Duckworth said earlier, I

1:16:20

hope that Intel's arc driver development

1:16:22

is improving Intel's integrated graphic, is

1:16:24

improving, integrated graphics, like they're the

1:16:26

same core, same core, same core,

1:16:29

same, same infrastructure. Now the question

1:16:31

is, is your laptop provider going

1:16:33

to give you the new driver

1:16:35

updates? Probably not, but you can

1:16:38

seek them out if you care,

1:16:40

is the important thing. Or maybe

1:16:42

Windows Update, or maybe Windows Update.

1:16:44

pipeline for good rasterized games, was

1:16:47

real good for a minute there

1:16:49

at like Windows Seven days and

1:16:51

now it's back to being, hey,

1:16:53

we're gonna give you something from

1:16:56

six months ago. Christine Decker has

1:16:58

one. How does this movement toward

1:17:00

AI FrameGen impact game developers? Are

1:17:02

they less responsible for good rasterized

1:17:05

games and able to lean on

1:17:07

DLSS to make it good? That

1:17:09

is actually a really good question

1:17:12

because it is I think the

1:17:14

opposite, right? Yeah, go on, you

1:17:16

start. I was just going to

1:17:18

say, all of the AI frame

1:17:21

gen and upscaling starts with a

1:17:23

rasterized frame. Or it may not

1:17:25

be a fully rasterized lighting. You

1:17:27

may use ray tracing to do

1:17:30

the lighting in that frame, whatever.

1:17:32

You start with one frame and

1:17:34

then it extrapolates from there. So

1:17:36

if you don't start with something

1:17:39

good, you're not garbage in, garbage

1:17:41

out, right? It's the normal rule.

1:17:43

Yeah. The normal rule applies. we're

1:17:45

seeing deficiencies I think are going

1:17:48

to be in things like frame

1:17:50

pacing and stuff like that for

1:17:52

games that don't support the frame

1:17:54

gen and all that is My

1:17:57

fear is that because you can

1:17:59

hide a lot of frame pacing

1:18:01

crimes with frame gen if your

1:18:03

game is performing enough That when

1:18:06

you run the same game on

1:18:08

AMD hardware, you're going to have

1:18:10

some real gnarly microstutter problems and

1:18:12

stuff like that and that's going

1:18:15

to be I think The irony

1:18:17

is for people on in video

1:18:19

cards, you're going to end up

1:18:21

having to do less testing, but

1:18:24

it's going to create a lot

1:18:26

more testing on Intel and AMD

1:18:28

hardware than most non-tribal A games

1:18:30

probably do today. So, I don't

1:18:33

know. Yeah, I, uh, how do

1:18:35

I phrase this? I don't like,

1:18:37

I hate the term, you know,

1:18:39

people, developers are using DLS as

1:18:42

a crutch. I hate that phrase.

1:18:44

It, like, to me, that belittles

1:18:46

the hard work that developers put

1:18:48

in to make these incredibly complex

1:18:51

games. But at the core of

1:18:53

this issue, that's how I feel

1:18:55

that... There have been a couple

1:18:57

of cases where games like didn't

1:19:00

run well as expected and they

1:19:02

said that's why we expect you

1:19:04

to turn on DLS. I don't

1:19:06

think that's acceptable. I think a

1:19:09

game should be able to run

1:19:11

well enough, hopefully on most people's

1:19:13

hardware because Intel, uh, because Intel

1:19:15

may have a large lead, but

1:19:18

it's not a console market. So

1:19:20

it matters. And again, I really

1:19:22

think... that even with DLS you

1:19:25

got to be putting out certain

1:19:27

numbers of quote unquote real frames

1:19:29

for it to still feel good.

1:19:31

So that's vitally important even if

1:19:34

you're using DLS. Adam asked a

1:19:36

question to Jensen Wong at CES,

1:19:38

asking, do you think there's ever

1:19:40

going to be a future where

1:19:43

you know games are entirely rented

1:19:45

by AI? And he said no,

1:19:47

because we still need the GPU

1:19:49

to render what is the ground

1:19:52

truth and how this thing should

1:19:54

look and then build off of

1:19:56

that. for those extra AI generated

1:19:58

frames. So. I think traditional

1:20:00

rasterization matters and I

1:20:03

hope that developers don't wind

1:20:05

up using DLS as a crutch because it's

1:20:07

not universal, not everyone loves

1:20:09

it. So on and so forth.

1:20:11

I think there's a thing that

1:20:13

people don't necessarily understand which is Like

1:20:15

the conversation between developers even small developers

1:20:18

like the the Devrel teams in invidia

1:20:20

to a huge invidious has a huge

1:20:22

Devrel team AMD and Intel to lesser

1:20:24

extent but when you when you're testing

1:20:27

something for example in the game that

1:20:29

I worked on last when we got

1:20:31

a complaint that people were having problems

1:20:33

with the specific era of AMD card

1:20:36

I reached out to our AMD rep I said

1:20:38

hey we're having problems on this and

1:20:40

this and this can you guys take

1:20:42

a look at it? and they passed

1:20:44

it off to their testing lab where

1:20:46

they generated an enormous report using debug

1:20:48

like they we gave them access to

1:20:50

the debug version of our game that

1:20:52

has the debug hooks so they could

1:20:54

hook all their instrumentation into it and

1:20:56

we got like a 50 page report

1:20:58

back from them that said okay on

1:21:00

this generation of hardware you have memory

1:21:02

management issues in these specific parts of

1:21:04

these levels and it's because of too

1:21:06

many draw calls here, too many lights

1:21:08

here, blah blah blah blah blah. And

1:21:10

like in video does the same thing,

1:21:12

Intel wasn't really relevant at the time we

1:21:14

were doing this. But you know, like when

1:21:16

we had a problem, when we were getting

1:21:19

ready to ship the game at the beginning,

1:21:21

our in video rep was like, hey, send

1:21:23

us a debug build and we'll let you

1:21:25

know how it runs on everything back to

1:21:28

like 20 80s and or 20 60s. And

1:21:30

it was an incredibly powerful thing. You still

1:21:32

have to triage those problems because you're like,

1:21:34

you know, you, am I going to go

1:21:37

out and do a 2050 mobile optimization? Probably

1:21:39

not. But, but when there was

1:21:41

a problem in the 2060s, yeah,

1:21:43

we absolutely spent time figuring that

1:21:45

out and got to a point

1:21:47

that like we weren't doing an

1:21:49

insane number of draw calls in

1:21:52

whatever section of the map

1:21:54

or whatever problems. Oh, by Elena,

1:21:56

thanks. Have fun answering to turning

1:21:58

all the cues into A's. I

1:22:00

guess that'll do it for

1:22:02

us this week then for

1:22:04

definitely Sign into the discord,

1:22:07

post your questions, we'll get,

1:22:09

if you have questions about

1:22:11

this in the future when

1:22:13

you're listening to this not

1:22:15

on the live feed, we'll

1:22:18

circle back next week and

1:22:20

hit some of those for

1:22:22

sure. Check back next week

1:22:24

for your Fix a PC

1:22:26

talk on the Full Nurd

1:22:29

to listen on the go,

1:22:31

subscribe to us on Apple

1:22:33

podcast, Spotify, Spotify already said.

1:22:35

And if you're on one

1:22:37

of those services, leave us

1:22:40

a review. Every time you

1:22:42

do, Brad buys a new

1:22:44

car. So congratulations, Brad. You're

1:22:46

going to be busy. No,

1:22:48

no, it's going to be

1:22:51

great. This is going to

1:22:53

be good for you. But

1:22:55

yeah, we'll see all next

1:22:57

week. Thanks for watching everybody

1:22:59

for, I guess, take us

1:23:02

out, Alex. All right, everybody.

1:23:04

Have a great week. Be

1:23:06

nice to each other. Take

1:23:08

it easy. Bye.