Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

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

 

[jpiniero]

It's good that AMD is continuing to support HPC... although I wonder how many researchers are still doing it? Figure most jumped on the AI bandwagon just because of the hype to get funding.

 

[adroc_thurston]

The Intel route?

Intel L1/SLM slab is the same as NV.

 

[maddie]

It's good that AMD is continuing to support HPC... although I wonder how many researchers are still doing it? Figure most jumped on the AI bandwagon just because of the hype to get funding.

Yep, I guess all work in nuclear fusion, advanced hypersonics weapons research, weather prediction, genetics, and thousands of other fields have stopped due to AI.

 

[Timorous]

Yep, I guess all work in nuclear fusion, advanced hypersonics weapons research, weather prediction, genetics, and thousands of other fields have stopped due to AI.

We are going to need nuclear fusion to power AI...

 

[maddie]

We are going to need nuclear fusion to power AI...

The rate it's going, we might need a Sun sized reactor.

Not that I'm complaining, but have others noticed the sudden "elite" embrace of nuclear and even fossil fuels to push the new paradise of AI. If I was a true believer in Anthropogenic GW, I would be feeling very betrayed right about now.

Sadly P&N never has sane & intelligent debates. Anyhow, enough of this here.

 

[eek2121]

AMD might respond to the super series. A little birdie told me that very slightly higher clocks (same chips, just better binning) + more VRAM (no GDDR7 though) would be in store for such a product. Oh and that little birdie is definitely not twitter. Also, no final decision has been made. 😉

 

[igor_kavinski]

The rate it's going, we might need a Sun sized reactor.

Interested to know why isn't anyone doing distributed AI? Like a AI@Home project for the good of humanity?

 

[igor_kavinski]

AMD might respond to the super series. A little birdie told me that very slightly higher clocks (same chips, just better binning) + more VRAM (no GDDR7 though) would be in store for such a product.

They have to. 18GB vs. 16GB makes them look bad from a value perspective. They gonna need at least 20GB.

 

[GodisanAtheist]

Interested to know why isn't anyone doing distributed AI? Like a AI@Home project for the good of humanity?

-At a guess the bandwidth requirements might be too extreme given the memory requirements for AI.

If you need dozens of gigs of local cache to train AI models the bandwidth requirements for passing that info around a DC network are probably off the charts and/or the speed is too slow for it to make any sense.

Skynet tho, so ???

 

[Kepler_L2]

Oh no it shouldn't even have LDS.

Looks like it's still there

 

[basix]

They have to. 18GB vs. 16GB makes them look bad from a value perspective. They gonna need at least 20GB.

In my opinion it does not really matter from a functional point of view. 16 GByte is enough for a 9070.

Sure, 18 looks bigger than 16 on paper. Practical effects are non-existent except for 1ppm of use cases. 12 vs. 16 could have noticeable effects with raytracing, high-res screens and textures. And also longevity of the card.

On the higher end of the spectrum (5070TiS / 5080S / 9070XTX) it might make sense to add some memory. Even though I still think it is of relatively little use on those cards.

 

[SolidQ]

GeForce have some regression

 

[Vikv1918]

GeForce have some regression

Nice. This means RDNA4 PPA and efficiency is now equal to Blackwell. This is kinda underrated what with all the focus on RT and AI. The last time AMD and nvidia were equal in PPA and efficiency on the same node was I think the R9 200 series 12 years ago.

 

[gdansk]

According to Micron GDDR6 uses 1.44x pj/bit compared to GDDR7. That's more of a "process" advantage than N4P.

 

[jpiniero]

According to Micron GDDR6 uses 1.44x pj/bit compared to GDDR7. That's more of a "process" advantage than N4P.

At similar speed though.

 

[gdansk]

At similar speed though.

Nope. At a higher speed.
It may not be that much more efficient when used in reality where less than peak bandwidth is needed from memory.

But it is a huge power advantage overall and saves shore area for GB205 too.

 

[SolidQ]

This means RDNA4 PPA and efficiency is now equal to Blackwell.

info from other source. would be interesting to see how much improvements RT/PT

 

[Paratus]

View attachment 126521
Nice. This means RDNA4 PPA and efficiency is now equal to Blackwell. This is kinda underrated what with all the focus on RT and AI. The last time AMD and nvidia were equal in PPA and efficiency on the same node was I think the R9 200 series 12 years ago.

Showed some nice gains at 1440p across multiple games. So definitely wasn’t just a single game or two bug fix.



Not quite so much at 4K but still not bad. Probably running into a different bottleneck. Maybe the memory subsystem?

 

[basix]

I suspect somewhat better bandwidth and infinity cache utilization. From another test it can be seen, that the 9070XT benefits more than a 9070. And some gains with pathtracing were incredible, up to +39%!

Radeon RX 9070 (XT) im Test: Höhere Performance dank Treiber

Radeon-Grafikkarten haftet der Ruf an, mit der Zeit besser zu werden - so auch RDNA 4 alias Radeon RX 9000. PCGH zeigt, was sich seit dem Start getan hat.

www.pcgameshardware.de

 

[tajoh111]

Some of the results in that review look suspicious. He admits in the 9600xt that Blackwell had gains in driver performance. Looks like they vanished for this review.

Spiderman remastered results look off when you look at their 9600xt review. He had same results as a RTX 5070.

Also the disappearance of war thunder because of game issues is suspicious since Nvidia had a 20 percent performance advantage here. Marvels rivals being dropped while having game issues while Nvidia had a advantage there is also suspect. When you combine the inconsistency in results, the game drops because of game "issues" and the regressions in performance, this seems like an AMD sponsored piece.

 

[gdansk]

Months ago I said HWUB would retest and they would have better results.

Because my similarly configured system was showing higher numbers than their initial review I am not convinced most of it is driver improvements (except where noted in AMD change logs). But this extra variation is what they will have with their manual test methods.

 

[GaiaHunter]

Marvel Rivals was apparently dropped due to replay file not being compatible.

If you go around and check reviews, there are several different performance numbers for all the games/cards.

For example on the 9070 XT release review:



Probably because he runs a new test/section.

 

[DAPUNISHER]

Some of the results in that review look suspicious.

Disagree. I'll explain below

Spiderman remastered results look off

He retested and rechecked 3 times. He is one of the most experienced and thorough reviewers with a solid methodology. No way he goofed that hard.

Also the disappearance of war thunder because of game issues is suspicious since Nvidia had a 20 percent performance advantage here.

There is nothing sus about it. My son is one of the OG War Thunder players. He and his squad are bent about the devs nerfing the game recently. He was telling me about it Saturday.

Marvels rivals being dropped while having game issues while Nvidia had a advantage there is also suspect.

Nothing sus here either. If you can't reproduce the exact same gameplay from a MP game, you can't say for certain if there are performance gains. This also was explained. No tinfoil needed.

When you combine the inconsistency in results, the game drops because of game "issues" and the regressions in performance, this seems like an AMD sponsored piece.

There it is, the shoot the messenger fallacy. AMD improving performance with drivers has so much precedence that they call it fine wine, as the video does. Accusing Steve of pimping AMD without disclosure of the sponsorship is entirely unwarranted and baseless.

 

[GaiaHunter]

So the 5070 TI and the 9070 XT traded blows and were essentially same tier GPUs and now the 5070 TI and the 9070 XT trade blows and are essentially same tier GPUs.


Absolute shills those HUB guys...

 

[tajoh111]

Disagree. I'll explain below

He retested and rechecked 3 times. He is one of the most experienced and thorough reviewers with a solid methodology. No way he goofed that hard.

There is nothing sus about it. My son is one of the OG War Thunder players. He and his squad are bent about the devs nerfing the game recently. He was telling me about it Saturday.

Nothing sus here either. If you can't reproduce the exact same gameplay from a MP game, you can't say for certain if there are performance gains. This also was explained. No tinfoil needed.

There it is, the shoot the messenger fallacy. AMD improving performance with drivers has so much precedence that they call it fine wine, as the video does. Accusing Steve of pimping AMD without disclosure of the sponsorship is entirely unwarranted and baseless.

You didn't address the counter strike issue. I will show you why the results are so suspect from their own data concretely. Below are the original 5070 review.

Here is the 9600xt review. You can't say different scene because of numbers from cards like the 4070 and 7800xt(they are the same, different scene would be different numbers). The 9060xt techspot review specifically mentions gains from blackwell.

These updated 5070 ti results are worse than a 5070 from their 9060xt review. If this isn't explicitly suspicious, I really can't help you. You cannot use the different scene excuse they mention the 479 fps in initial review.

And if you are going to throw our results because they are buggy, why did they keep the spiderman ones? The RTX 5070 ti numbers are the same as an RTX 5070.

More issues can be seen in this if we cross reference the 9600xt review with same the original RTX 5070 review.

Look at cyberpunks results. Launch results with a RTX 5070.

Now look at the 9060xt review with updated 5070 results.

There is a substantial gain of 12%. More than the gains from updated 9070 results(7%). Considering how much faster the 5070 ti is than the RTX 5070 ti, the results make no sense.

Stalker also shows this. Launch review for the RTX 5070 which has the same results in this updated comparison for the RTX 5070 ti(so no different scene excuse).

Now here are the results for the 5070 in the 9060xt review. Again significant gains from driver updates from 52 to 58 FPS. you can't use the difference scene excuse the numbers for the 7800xt and 4070 are the same in both reviews. Where are these gains for the RTX 5070 ti in this comparison? Do you think the RTX 5070 ti is only 10% faster than a 5070 is an accurate result?

They are only testing two cards, buggy results is a half ass excuse to not include 2 popular games still played now. Warthunder and marvels rival have around 50k active players today compared to games to starfield which has 2000 players or even hogwarts legacy which has around 10k players.

Too many weird results affecting too high a percentage of the sample size of games. If Steve is not being paid by AMD, he should have retested with a different system because too many buggy results to accept this comparison as valid.