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

 

[techjunkie123]

https://videocardz.com/pixel/amd-navi-48-rdna4-gpu-has-53-9-billion-transistors-more-than-nvidia-gb203

Die size confirmed to be 357 mm^2. So density is indeed 150 MTr/mm^2. And 64 MB of infinity cache.

 

[SolidQ]

Final specs

 

[branch_suggestion]

https://videocardz.com/pixel/amd-navi-48-rdna4-gpu-has-53-9-billion-transistors-more-than-nvidia-gb203

Die size confirmed to be 357 mm^2. So density is indeed 150 MTr/mm^2. And 64 MB of infinity cache.

Final specs

Lots of big revelations there, like WMMA is twice the raw throughput which I was really hoping for.
So ~800TF FP8 using SWMMAC instructions, which is on par with AD103.
Oh, the ROPS line got nuked, hmm... (and it's 128, nothingburger, though please do update the pixel fill already)

 

[SolidQ]

interesting NV is calculate for INT4?

 

[CastleBravo]

Final specs

Non-XT might OC to near XT performance if the power limit can be raised high enough, especially if vram bandwidth is the main limitation.

 

[branch_suggestion]

interesting NV is calculate for INT4?

NV number is FP4 with sparsity.
N48 is roughly equal in FP8 to AD103, not far behind full GB203.

 

[SolidQ]

NV number is FP4 with sparsity.

Yeah found this

RTX 5070 AI TOPS VS 4080

The AI TOPS of RTX 5XXX are FP4 AI TOPS, the AI TOPS for RTX 4XXX are FP8 AI TOPS, i dont know if TL does generate images in FP8, but clearly not in FP4 at the moment. AI TOPS besides FP4 are not that different for RTX 4XXX vs 5XXX. RTX 4XXX cant do FP4. Rumors are for RTX 5090 = 40% faster...

community.topazlabs.com

N48 is roughly equal in FP8 to AD103, not far behind full GB203.

That mean beat 5070ti in FP4? 1405 vs 1557 RDNA4 if i right understand

 

[Heartbreaker]

https://videocardz.com/pixel/amd-navi-48-rdna4-gpu-has-53-9-billion-transistors-more-than-nvidia-gb203

Die size confirmed to be 357 mm^2. So density is indeed 150 MTr/mm^2. And 64 MB of infinity cache.

Confirmed where. That link says it's a leak. It's old news we were discussing days ago when it came out. It's source is a Chinese forum post that says:

According to the personal test of a guy who has the most profound hardware knowledge in this forum, it is about 350mm², which is similar to the result of strict measurement and calculation by someone in the anandtech forum based on official photos.

Going to Videocardz, there is this newer story, maybe you meant to post this link:

https://videocardz.com/newz/final-specifications-of-amd-radeon-rx-9070-xt-and-rx-9070-gpus-leaked

 

[branch_suggestion]

Non-XT might OC to near XT performance if the power limit can be raised high enough, especially if vram bandwidth is the main limitation.

Vega 56 all over again, but AMD will probably stop the poors from having fun.

Yeah found this

RTX 5070 AI TOPS VS 4080

The AI TOPS of RTX 5XXX are FP4 AI TOPS, the AI TOPS for RTX 4XXX are FP8 AI TOPS, i dont know if TL does generate images in FP8, but clearly not in FP4 at the moment. AI TOPS besides FP4 are not that different for RTX 4XXX vs 5XXX. RTX 4XXX cant do FP4. Rumors are for RTX 5090 = 40% faster...

community.topazlabs.com

That mean beat 5070ti in FP4? 1405 vs 1557 RDNA4 if i right understand

N48 does not support FP4, only INT4.

 

[SolidQ]

Vega 56 all over again, but AMD will probably stop the poors from having fun.

But also XT can be OC'ed. We still need tests how many OC add % and PowerCons

 

[branch_suggestion]

But also XT can be OC'ed. We still need tests how many OC add % and PowerCons

Memory oc is what truly matters.
+500 is a 20% bump assuming it is stable.
But yeah, with GDDR7 and identical power to GB203 it should...

 

[SolidQ]

What is interesting it's pixel fillrate is reduced, even vs 7800XT

 

[branch_suggestion]

What is interesting it's pixel fillrate is reduced, even vs 7800XT

Which is fascinating, but it really implies >2 SA configs in the future per SE.

 

[branch_suggestion]

N48 scales better at high res than N31 assuming no memory bottleneck.
One sample was hitting 3.35Ghz in games at 350-360W with roughly 6-8% more perf.

 

[SolidQ]

Now question how many ROPS
UPD:
Show 64, i don't see now them in news. Seems they remove it

One sample was hitting 3.35Ghz in games at 350-360W with roughly 6-8% more perf.

6-8% not very much, i was think about 10-12%

 

[leoneazzurro]

Now question how many ROPS


6-8% not very much, i was think about 10-12%

BW limitations, probably.

 

[branch_suggestion]

6-8% not very much, i was think about 10-12%

Still enough to be nipping the 5080's heels with far less membw.
Now if you double the SEs, quadruple the SA/shaders, quadruple the L3, triple the membw...

 

[gaav87]

Still enough to be nipping the 5080's heels with far less membw.
Now if you double the SEs, quadruple the SA/shaders, quadruple the L3, triple the membw...

Guess 1080p perf suck then womp womp
Also my guess mem OC gives way more perf uplift than core oc.

 

[SolidQ]

Now if you double the SEs, quadruple the SA/shaders, quadruple the L3, triple the membw...

yeah but that killed
Interesting if RDNA5 is again new arch, maybe xx7x class gonna be 4090 level?

 

[branch_suggestion]

Guess 1080p perf suck then womp womp

That isn't the point, the uArch scales better with less resources at higher resolutions that favour higher resources by default.
It is super hungry for things to execute in a more efficient manner.

 

[branch_suggestion]

yeah but that killed
Interesting if RDNA5 is again new arch, maybe xx7x class gonna be 4090 level?

Well I'm implying what could still be done for GFX13, the 2023 roadmap that was leaked is not gospel.

 

[MrTeal]

All for it, as long as it isn't the Voodoo of 3dfx kind ...

If AMD gives me a 400W card that supports stacking and high degrees of SSAA instead of TAA I would buy 4 of em. I can tuck a couple MORAs somewhere.

 

[coercitiv]

If AMD gives me a 400W card that supports stacking and high degrees of SSAA instead of TAA I would buy 4 of em.

That would be the dream, 4 cards synchronously rendering the game at slightly different angles to obtain pure AA. No jitter, no blur, just1600W worth of smoothness.

 

[gaav87]

Now question how many ROPS

All of them

 

[Josh128]

They cared about NVENC before.
The goalposts will move towards whatever thing AMD doesn't have/is worse at.

It always falls back to "drivers". For some reason, a large number of people still believe Radeon drivers are worse than Nvidias. I dont know if AMD will ever be free of that stigma.