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

 

[gaav87]

Does 16x32 mean doubled INT4 rate for RDNA4 compared to RDNA3?

For the V_SWMMAC_I32_16X16X64_IU4 (rdna4 swmma)
INT4
A: 16×16 stored/32×16 actual
B: 16×64
Result: 32×64 INT32

For the V_WMMA_I32_16X16X32_IU4: (rdna4 new wmma)
A: 16×16 INT4
B: 16×32 INT4
Result: 16×32 INT32

For the V_WMMA_I32_16X16X16_IU4: (rdna3 wmma)
iNT4
Result: 16x16 INT32

Dimension:
16×16 INT32 matrix: 256 elem × 4 bytes= 1 KB
32×64 INT32 matrix: 2048 elem × 4 bytes= 8 KB
So rdna4 swmma int32 matrix is 8 time larger than rdna3
and wmma 2x as big as 16x16

 

[gaav87]

Does 16x32 mean doubled INT4 rate for RDNA4 compared to RDNA3?

So yes the 16×32 INT4 = higher rate of computation compared to 16×16 INT4 theorethicaly 16×32 INT4 operation can process 2x as many elements in parallel vs rdna3. ofc if the mem bandwidth and processing are not bottlenecking.

 

[gaav87]

INT4 WMMA exists already on RDNA3. The only question is its rate. Same as INT8/FP8 or double.
https://gpuopen.com/learn/wmma_on_rdna3/
https://www.amd.com/content/dam/amd...r-instruction-set-architecture-feb-2023_0.pdf -> chapter 7.9

But afik FP8 does not exist on rdna3 and it does on rdna4.
FP8 result in 16x16 FP32 on rdna4 on wmma and 16x32 fp32 on swmma

 

[basix]

Yeah, chipsandcheese did provide some neat overview of the RDNA4 ISA changes: https://chipsandcheese.com/p/examining-amds-rdna-4-changes-in-llvm

 

[gaav87]

Yeah, chipsandcheese did provide some neat overview of the RDNA4 ISA changes: https://chipsandcheese.com/p/examining-amds-rdna-4-changes-in-llvm

User Guide for AMDGPU Backend — LLVM 20.0.0git documentation

llvm.org

They had swmma in here but i can no longer find it or im too tired anyway.
gfx12=rdna4 if u dont folow the leaks etc.

 

[Kepler_L2]

Does 16x32 mean doubled INT4 rate for RDNA4 compared to RDNA3?

RDNA4 doubles all rates for DP4a/WMMA.

 

[Kepler_L2]

Not here either.

Yeah it's using FP8 on RDNA4.

 

[gaav87]

Yeah it's using FP8 on RDNA4.

Yes fp8 but dont know if in future it will not be used in something and fp16 on older. (edit: i mean int4)

 

[basix]

RDNA4 doubles all rates for DP4a/WMMA.

I assume that doubling covers only <=8bit datatypes and not 16bit as well?

 

[Kepler_L2]

I assume that doubling covers only <=8bit datatypes and not 16bit as well?

16 bit too.

 

[Gideon]

AMD released some dibits of info:

Loading…

videocardz.com

• Optimized Compute Units
• Enhanced AI Compute capabilities (2nd-generation AI accelerators)
• Improved Raytracing performance per Compute Unit (3rd-generation Raytracing accelerators)
• Better media encoding quality (2nd-generation AMD Radiance Display Engine)
• 4nm manufacturing process

But no performance slides. I guess the GPUs are still quite a while away, or the comparison is not that flattering to RDNA4. Might also just wait for Nvidia to make the next move. Either way this probably means we still have to wait quite a bit until we get more info

 

[jpiniero]

Probally more letting nVidia dictate what the price is.

 

[coercitiv]

But no performance slides. I guess the GPUs are still quite a while away, or the comparison is not that flattering to RDNA4. Might also just wait for Nvidia to make the next move. Either way this probably means we still have to wait quite a bit until we get more info

It's almost like they need permission from Nvidia to speak. First they have to wait and see what Nvidia does, then they come out barking at the fence for 15 minutes during a keynote. Rinse and repeat in 2 years.

 

[adroc_thurston]

It's almost like they need permission from Nvidia to speak. First they have to wait and see what Nvidia does, then they come out barking at the fence for 15 minutes during a keynote. Rinse and repeat in 2 years.

No halo means you can't talk. Simple as.
They can't command the pricing stack without a 3kilobick tiled monstrosity. Which they killed.

 

[jpiniero]

No halo means you can't talk. Simple as.
They can't command the pricing stack without a 3kilobick tiled monstrosity. Which they killed.

AMD's Fire Sales are a bigger problem than anything nVidia does.

 

[gaav87]

Also i just heard all games that have fsr3.1 will work on fsr4

Well this just got confirmed

 

[adroc_thurston]

AMD's Fire Sales are a bigger problem than anything nVidia does.

They're actually normal. It's NV not dropping the street price that is the outlier.

Well this just got confirmed
View attachment 114304

Yeah, it's very PnP.

 

[xnery007]

Did they seriously announce that they’ll announce in the future what the cards will even be? Alright then, good thing we have the biggest tech conference to show off some mockups.

 

[jpiniero]

They're actually normal. It's NV not dropping the street price that is the outlier.

People buy the Fire Sale cards instead of buying the new ones.

 

[poke01]

Yawn, what a boring and bad announcement.

 

[poke01]

They're actually normal. It's NV not dropping the street price that is the outlier.

Companies with mindshare don’t drop prices very quickly

 

[adroc_thurston]

People buy the Fire Sale cards instead of buying the new ones.

Hell no, AMD dGPUs share is pitiful.

Yawn, what a boring and bad announcement.

They can't do anything without the chainsaw. I said that and they very much know that too.

 

[adroc_thurston]

Companies with mindshare don’t drop prices very quickly

Yeah they do, g92 rebrands were selling for pennies despite le g80 mindshare. The goys are too indoctrinated now I guess.

 

[marees]

Well this just got confirmed
View attachment 114304

Why does it say "only available on 9070 series" ??

 

[jpiniero]

Why does it say "only available on 9070 series" ??

Looks like they took a chainsaw to N44 to get the cost down.