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

 

[adroc_thurston]

Maybe AMD should bring back Raja to lead the Radeon software side. Raja for all his faults related to hardware, did reasonably well in software side

Radeon was also losing money under him. Let's not.

 

[yuri69]

Maybe AMD should bring back Raja to lead the Radeon software side. Raja for all his faults related to hardware, did reasonably well in software side. The were lots of improvements in software during his time. BTW, what has happened to periodic Adrenal releases? It got lost after Rajas exit.

Raja also did:
* bring stupid marketing shows like AMD Capsaicin in
* present hilariously cherry picked "up to" numbers for Polaris 2.8x PPW
* stupidly hype the delayed Vega just to release the "Frontier Edition" to meet the H1 2027 release
* twist the "Frontier Edition vs regular 64 edition" gaming speed difference
* go with the "Poor Volta" campaign
* not confirm there were no magic drivers enabled Vega's magic stuff like DSBR or HBCC

I don't know if things such like that would be really worth it compared to any SW improvements.

 

[gaav87]

Who said that even.

They really don't want to? The only way to challenge the comp in client dGFX is the Chainsaw.

You're going right back to at least 3 dies with RDNA5. Hope that helps!

Chainsaw is real. Just look at 9800x3d half the MT of 285k for same street price 15% faster in gaming sells like hot cakes.

 

[adroc_thurston]

Chainsaw is real. Just look at 9800x3d half the MT of 285k for same street price 15% faster in gaming sells like hot cakes.

nT AMD has a leadership product too. That's not even touching the threadrippers.

 

[gaav87]

4 for 4 for 4

 

[Saylick]

4 for 4 for 4
View attachment 114182

FSR4 supported only on RDNA4…

 

[linkgoron]

Maybe AMD should bring back Raja to lead the Radeon software side. Raja for all his faults related to hardware, did reasonably well in software side. The were lots of improvements in software during his time. BTW, what has happened to periodic Adrenal releases? It got lost after Rajas exit.

Yeah, we saw how successful he was with the software at Intel.

Also, he was awful. There's a reason why he's no longer at Intel as well.

 

[RnR_au]

Looks like the video link is already up...

 

[lucasworais]

Yeah, i wonder why they didn't make a 80CU RDNA4 monolith size

Or just glue two n48 together

 

[PJVol]

Or just glue two n48 together

Indeed, 120 (or what, 128? 8sax16) CU's would stomp all over 5080's at the bare minimum )
It's strange they still hasn't reach 8sa limit despite being allowed by architecture itself since rdna2 (?)

 

[JustViewing]

Raja also did:
* bring stupid marketing shows like AMD Capsaicin in
* present hilariously cherry picked "up to" numbers for Polaris 2.8x PPW
* stupidly hype the delayed Vega just to release the "Frontier Edition" to meet the H1 2027 release
* twist the "Frontier Edition vs regular 64 edition" gaming speed difference
* go with the "Poor Volta" campaign
* not confirm there were no magic drivers enabled Vega's magic stuff like DSBR or HBCC

I don't know if things such like that would be really worth it compared to any SW improvements.

I completely agree. However after he left, the Radeon group went into a shell. There are no enthusiasm or false promises.

 

[linkgoron]

I completely agree. However after he left, the Radeon group went into a shell. There are no enthusiasm or false promises.

The good old days of hyping products, overpromising and underdelivering. I much prefer todays AMD that just underdelivers.

 

[JustViewing]

I much prefer todays AMD that just underdelivers.

 

[Win2012R2]

I much prefer todays AMD that just underdelivers.

They can't underdeliver if they had not promised anything. But they can still meet or even exceed the expectations...

 

[beginner99]

A tir

They can't underdeliver if they had not promised anything. But they can still meet or even exceed the expectations...

A third iteration of 6800 xt performance for almost same price is hardly not under-delivering. but same is true for NV so...

 

[gdansk]

A third iteration of 6800 xt performance for almost same price is hardly not under-delivering. but same is true for NV so...

This time it should be 1.2x 6800 XT at a lower MSRP. But we'll see.

 

[Kepler_L2]

FSR4 supported only on RDNA4…

It's not, though RDNA4 has an exclusive kernel.

 

[gaav87]

It's not, though RDNA4 has an exclusive kernel.

Exclusive kernel sounds like its only for rdna4 xD

 

[Saylick]

It's not, though RDNA4 has an exclusive kernel.

That’s good to hear. Any idea if competitors can use it as well?

 

[Kepler_L2]

Exclusive kernel sounds like its only for rdna4 xD

No it's like XeSS, there are different models.

 

[adroc_thurston]

This time it should be 1.2x 6800 XT at a lower MSRP. But we'll see.

Lower power too.

 

[Kepler_L2]

That’s good to hear. Any idea if competitors can use it as well?

Depends on how they are implementing it, if it's DirectSR I think all the FSR4 models will be included and selected by the driver.

 

[gaav87]

Depends on how they are implementing it, if it's DirectSR I think all the FSR4 models will be included and selected by the driver.

Well i heard its not like xess

 

[SolidQ]

No it's like XeSS, there are different models.

RDNA2 via DP4a?

 

[basix]

From what I heard FSR4 uses FP8 path on RDNA4 and what I speculate FP16 path on everything else (older AMD GPUs, Intel, Nvidia, Consoles, APUs). Maybe there is some additional "Block-FP16 = FP8 Throughput" path for AMD APUs with XDNA2 NPU (Strix Point & Halo).

FP8 and FP16 makes sense to me. It is more accurate than INT8 (=less DNN parameters required or higher DNN quality) and INT8 is a mess on AMD GPUs (e.g. N10 has lower INT8 rate than its smaller siblings, PS5 is unclear, RDNA3 has same INT8 rate as FP16, etc.) and would get killed by Nvidia Tensor Core INT8 (they simply have much higher throughput). FP8 throughput on RDNA4 should be on a similar level like FP16 Tensor on similarly sized Nvidia and Intel GPUs. With INT8 it would probably be slower on RDNA2/3/4 compared to their Nvidia counterparts. Not a good idea.

Well i heard its not like xess

It seems to me to be more DLSS alike or what ARM does on mobile (see their Siggraph presentation from 2024). Parameter prediction, the rest is very similar to FSR2/3. XeSS seems to rely on a much heavier DNN.

FSR4 precursor:
https://gpuopen.com/learn/neural_supersampling_and_denoising_for_real-time_path_tracing/