Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.

Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

gaav87 · Jan 9, 2025

SolidQ said:
chinese guy, who leak to us 4080 -5%, was surprised by speedway results were too low, despite in games like ~4080 RT

Ye its litteraly 2x perf of rx6800 (non-xt) i have no words.

iLLusiveMan · Jan 9, 2025

gaav87 said:
Well i think the TSE and SPEEDWAY chinese leaks are fake...
7900xt right there... Scores fall in line with top30-40 7900xt oc

View attachment 114618

Hype train AyyMDead

Hitman928 · Jan 9, 2025

gdansk said:
I won't believe anything until I see it.

SolidQ · Jan 9, 2025

- It's an internally developed demo using Unreal Engine.

- Performance was around 35 FPS,

Not representative, but similar walking room

blckgrffn · Jan 9, 2025

gdansk said:
I won't believe anything until I see it.

What if it’s a neural rendered fantasy? I mean, is anything you see real or just an approximation of reality that our brain considers good enough to function?

Are those shadows on the cave wall the real ones or are they being upscaled and AI refined into something else? How can we ever know the truth?!?

Haha, man I loved Intro to Philosophy.

soresu · Jan 9, 2025

blckgrffn said:
Are those shadows on the cave wall the real ones or are they being upscaled and AI refined into something else? How can we ever know the truth?!?

Obviously you can't reconstruct detail that isn't there.

But you can infer it from game engine data that a scaler based on rendered pixels and their temporal correlation alone would not be privy too.

I don't know enough about how ray reconstruction works to comment on it, but it sounds like the same thing applied to ray casting data being fed into the scaler to increase the accuracy and/or consistency of the output frames.

You do this by rendering a whole boat load of higher resolution frames across a wide range of content and use this as the 'ground truth' training data for the AI model to reference so that the weights are aligned to the correct inference from each new input.

soresu · Jan 9, 2025

adroc_thurston said:
The roadmap is out there, big dawg.

Is that the blurry roadmap photo posted a few days back?

soresu · Jan 9, 2025

SolidQ said:
Not representative, but similar walking room

Honestly the path tracing perf relative to Lumen is super impressive.

blckgrffn · Jan 9, 2025

blckgrffn said:
What if it’s a neural rendered fantasy? I mean, is anything you see real or just an approximation of reality that our brain considers good enough to function?

Are those shadows on the cave wall the real ones or are they being upscaled and AI refined into something else? How can we ever know the truth?!?

Haha, man I loved Intro to Philosophy.

soresu said:
Obviously you can't reconstruct detail that isn't there.

But you can infer it from game engine data that a scaler based on rendered pixels and their temporal correlation alone would not be privy too.

I don't know enough about how ray reconstruction works to comment on it, but it sounds like the same thing applied to ray casting data being fed into the scaler to increase the accuracy and/or consistency of the output frames.

You do this by rendering a whole boat load of higher resolution frames across a wide range of content and use this as the 'ground truth' training data for the AI model to reference so that the weights are aligned to the correct inference from each new input.

It’s a philosophical thought exercise reference.

An Allegory. A somewhat purposefully obscure joke. To show off how cultured and smart I am. Obligatory /s!

Allegory of the cave - Wikipedia

en.m.wikipedia.org

The point being, to be annoying I suppose, is that at best all our senses and perceptions are inference, so arguing about real pixels is silly in its own way.

But also, how will we know the real pixels from the fakes ones?!? Does it even matter?!? So on and so forth.

gdansk · Jan 10, 2025

blckgrffn said:
The point being, to be annoying I suppose, is that at best all our senses and perceptions are inference, so arguing about real pixels is silly in its own way.

I know you are having fun but the machine inference may mess with my senses. When an apparent image is wrong, sometimes obviously so, it can mess with my perception and break immersion in an annoying way.

When some pixels are part of an approximated frame inserted to increase the perception of motion, but it instead weakens the apparent motion, it is a failed fake. Or at least erroneous and counterproductive and against the purpose of the feature.

beginner99 · Jan 10, 2025

marees said:
"One of the weirder GPU launches"

just from the naming we can see how AMD planned to position and price these cards. I suspect they had some info on 5000 series performance but that info was wrong as in including fake frames. They realized that rather late and with that also that they can change (increase) the price of their product.

SolidQ · Jan 10, 2025

got from other source

1. They decided to not mess it up this time and rather observe Nvidia and respond more appropriately, so the actual reveal is set for a later date

2. He says that most gamers these days don't want to spend $1000 in a GPU, they released $1000 GPUs last gen but only got a very favorable response from the market when they lowered the prices and released the 7900 GRE.

3. The new GPUs won't cost $300 but they won't cost $1000 either, because 90% of gaming market is anywhere in between

4. FSR 4 requires lots of Machine Learning power, that's the reason why it's RDNA 4 exclusive, but they're making efforts to optimize it as much as possible so """"maybe"""" they can put it in RDNA 3 too.

5. They understand that a ML solution has much better quality, but they won't stop investing in FSR 3, it's still continuing to evolve.

6. RDNA 4 has "much much better, but muuuuuch better" Raytracing performance, he insists saying it twice in the interview

7. Developing a GPU chip is expensive in part because of how many chips have to be developed and that the prices of all of them spread accros all the model prices, so they rather focused on a more limited selection to keep prices as low as possible in the range that most gamers put their money on

8. He says that AMD have been learning a lot from their mistakes on the GPU segment

9. There are comments about the new CPU and the new mobile chip but they're not too relevant IMO

Bryo4321 · Jan 10, 2025

SolidQ said:
got from other source

4. FSR 4 requires lots of Machine Learning power, that's the reason why it's RDNA 4 exclusive, but they're making efforts to optimize it as much as possible so """"maybe"""" they can put it in RDNA 3 too.

I hope they can realize this on RDNA3, impressions of fsr4 seem promising. And hopefully next gen branch back into the higher end and give us a $999 card for high refresh 4K.

Also, hello gaf pal! I’ve seen your name before.

SolidQ · Jan 10, 2025

Bryo4321 said:
give us a $999 card for high refresh 4K.

That their plan max 1k for UDNA

Also, hello gaf pal! I’ve seen your name before.

Saylick · Jan 10, 2025

SolidQ said:
got from other source

1. They decided to not mess it up this time and rather observe Nvidia and respond more appropriately, so the actual reveal is set for a later date

2. He says that most gamers these days don't want to spend $1000 in a GPU, they released $1000 GPUs last gen but only got a very favorable response from the market when they lowered the prices and released the 7900 GRE.

3. The new GPUs won't cost $300 but they won't cost $1000 either, because 90% of gaming market is anywhere in between

4. FSR 4 requires lots of Machine Learning power, that's the reason why it's RDNA 4 exclusive, but they're making efforts to optimize it as much as possible so """"maybe"""" they can put it in RDNA 3 too.

5. They understand that a ML solution has much better quality, but they won't stop investing in FSR 3, it's still continuing to evolve.

6. RDNA 4 has "much much better, but muuuuuch better" Raytracing performance, he insists saying it twice in the interview

7. Developing a GPU chip is expensive in part because of how many chips have to be developed and that the prices of all of them spread accros all the model prices, so they rather focused on a more limited selection to keep prices as low as possible in the range that most gamers put their money on

8. He says that AMD have been learning a lot from their mistakes on the GPU segment

9. There are comments about the new CPU and the new mobile chip but they're not too relevant IMO

I didn’t know Frank spoke fluent Spanish, but it’s good to hear him speak more comfortably in this interview than that which I’ve seen before.

gdansk · Jan 10, 2025

beginner99 said:
They realized that rather late and with that also that they can change (increase) the price of their product.

It would be funny and backward if that's true.

1. People want AMD to have aggressive pricing so they can buy Nvidia at lower prices.
2. AMD knows people want to buy Nvidia at a lower price, so AMD plans to increase their pricing (even if it ruined their CES) such that Nvidia won't bother to decrease prices.

Is this the real jebait?

coercitiv · Jan 10, 2025

gdansk said:
It would be funny and backward if that's true.

1. People want AMD to have aggressive pricing so they can buy Nvidia at lower prices.
2. AMD knows people want to buy Nvidia at a lower price, so AMD plans to increase their pricing (even if it ruined their CES) such that Nvidia won't bother to decrease prices.

This is the strategy they used with RDNA3 at the high-end, and it also backfired. They can't track and sell "just under" Nvidia, they either come out with a good price for what they're offering or the launch goes badly. With RDNA3 they knew they messed up even against RDNA2, they should have priced accordingly.

If N48 is that much better than their existing lineup and FSR4 is pumping on all cylinders, then price it like 6800XT. However, if the performance jump is incremental and FSR4 still needs time to reach maturity (and more games), then it's 7800XT pricing. They either embrace what they have from day one or the market will force the conclusion upon them in a month or two, minus the positive media coverage.

adroc_thurston · Jan 10, 2025

beginner99 said:
just from the naming we can see how AMD planned to position and price these cards. I suspect they had some info on 5000 series performance but that info was wrong as in including fake frames. They realized that rather late and with that also that they can change (increase) the price of their product.

something about blackwell in general seems off.

SolidQ said:
FSR 4 requires lots of Machine Learning power, that's the reason why it's RDNA 4 exclusive, but they're making efforts to optimize it as much as possible so """"maybe"""" they can put it in RDNA 3 too.

wonder if it's FP8 aka numerics-limited

gdansk · Jan 10, 2025

coercitiv said:
This is the strategy they used with RDNA3, and it also backfired.

It isn't something I consider a good idea or likely.
But some of RDNA3 pricing problem was because Nvidia did cut the price of the 4080 12GB. It was supposed to be the same price as the 7900 XT.

Shmee · Jan 10, 2025

I don't know if the 3x8 pin power connectors will help, and I kind of doubt it. Though they helped with 7900XTX (many high end card designs had 3 and had better boost and overall performance due to more power) but with RDNA4 die being smaller, I doubt this will be the same case. But who knows, will have to see how the cards actually do in tests and how they actually sell for.

Thunder 57 · Jan 10, 2025

Shmee said:
I don't know if the 3x8 pin power connectors will help, and I kind of doubt it. Though they helped with 7900XTX (many high end card designs had 3 and had better boost and overall performance due to more power) but with RDNA4 die being smaller, I doubt this will be the same case. But who knows, will have to see how the cards actually do in tests and how they actually sell for.

It just occurred to me I've never had a GPU that used more than 200W. To be fair, these high watt cards seem to be a fairly new phenomenon. Three 8 pin cables just seems crazy to me.

soresu · Jan 10, 2025

beginner99 said:
They realized that rather late and with that also that they can change (increase) the price of their product

Given the dramatic success of Intel's B580 that would be a mistake.

The desire is there, the price is not, and AMD are not exactly swimming in market share at the moment.

soresu · Jan 10, 2025

Thunder 57 said:
It just occurred to me I've never had a GPU that used more than 200W. To be fair, these high watt cards seem to be a fairly new phenomenon. Three 8 pin cables just seems crazy to me.

Same, I've tried to stick to <200W until now, but it seems like that is increasingly becoming untenable.

Thunder 57 · Jan 10, 2025

soresu said:
Same, I've tried to stick to <200W until now, but it seems like that is increasingly becoming untenable.

Living in a warmer climate I prefer lower watts to avoid dumping heat into any one room. I'm not sure what my next upgrade will be but I have a feeling it will be closer to the 250W area. Luckily I don't typically game in marathon sessions so it hasn't been a problem.

soresu · Jan 10, 2025

adroc_thurston said:
something about blackwell in general seems off.

I'd be the first to moan about RDNA3's misleading dual issue FLOPS figures, but nVidia have been pushing increasingly silly numbers over the last several generations.

We often take a pot shot at tensor cores taking up room on nVidia's GPUs, but I'm beginning to wonder if they are actually cannibalising the general compute/gfx/CUDA cores in order to make room for more/beefier tensor cores to push the AI craze until the bubble explodes.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Senior member

Member

Diamond Member

Golden Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Golden Member

Member

Golden Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Memory & Storage, Graphics Cards Mod Elite Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member

Diamond Member