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

Mopetar · Saturday at 12:01 PM

Assembly work means knowing a little about what you're going to assemble. Even if it isn't full specs, you'll still have timelines and an idea of how much work you'll be doing. From that a person can make reasonable inferences.

jpiniero · Saturday at 12:05 PM

Hail The Brain Slug said:
Because of this, I'm highly skeptical of the legitimacy or accuracy of the "leaked roadmap", but here it is - freshly circulating today.

Agree. But it's realistic at least.

Joe NYC · Saturday at 12:19 PM

Hail The Brain Slug said:
Because of this, I'm highly skeptical of the legitimacy or accuracy of the "leaked roadmap", but here it is - freshly circulating today.

View attachment 101670

It will be interesting to see if Medusa Halo is going to be RDNA5

maddie · Saturday at 1:12 PM

Hail The Brain Slug said:
Because of this, I'm highly skeptical of the legitimacy or accuracy of the "leaked roadmap", but here it is - freshly circulating today.

If true, then no core chiplets in RDNA 5.

SolidQ · Saturday at 2:09 PM

Interesting, maybe in one day AMD will buyback Dresden GlobalFoundries

https://twitter.com/x/status/1803867395070857447

lucasworais · Saturday at 4:54 PM

Hail The Brain Slug said:
Because of this, I'm highly skeptical of the legitimacy or accuracy of the "leaked roadmap", but here it is - freshly circulating today.

View attachment 101670

So no halo RDNA5

jpiniero · Saturday at 5:09 PM

maddie said:
If true, then no core chiplets in RDNA 5.

Good catch. So def fake.

gaav87 · Saturday at 7:15 PM

"aco: use 1.5x vgprs for gfx1151 and gfx12"
FeatureMaxHardClauseLength32,
Feature1_5xVGPRs]>;

if (family == CHIP_NAVI31 || family == CHIP_NAVI32 || family == CHIP_GFX1151 ||
gfx_level >= GFX12) {
program->dev.physical_vgprs = program->wave_size == 32 ? 1536 : 768;
program->dev.vgpr_alloc_granule = program->wave_size == 32 ? 24 : 12;
} else {
program->dev.physical_vgprs = program->wave_size == 32 ? 1024 : 512;
if (gfx_level >= GFX10_3)

if (program->gfx_level >= GFX12) {
/* Same as GFX11, except one less for VSAMPLE. */
program->dev.max_nsa_vgprs = 3;

In the code, gfx12 VGPR configuration:
- If the wave size is 32, gfx12 has 1536 physical VGPRs.
- If the wave size is 64, gfx12 has 768 physical VGPRs.
The VGPR allocation granule for gfx12 is:
-24 for a wave size of 32.
-12 for a wave size of 64.
7900xtx has 1536 physical vgprs per simd does this mean rdna4 will have 2304 vgpr per simd as the feature has not been merged to LLVM ?

inquiss · Saturday at 10:01 PM

Joe NYC said:
It will be interesting to see if Medusa Halo is going to be RDNA5

Think it will be. Didn't adroc say that the RDNA 4 igpus were canned in favour of going straight to RDNA5

blckgrffn · Saturday at 11:51 PM

inquiss said:
Think it will be. Didn't adroc say that the RDNA 4 igpus were canned in favour of going straight to RDNA5

I think to be clear he said RDNA4 iGPU was axed in his his usually succinct manner. That would logically lead to RDNA5 being the next iGPU target unless there is a RDNA4.5 in the wings?

inquiss · Sunday at 7:35 AM

blckgrffn said:
I think to be clear he said RDNA4 iGPU was axed in his his usually succinct manner. That would logically lead to RDNA5 being the next iGPU target unless there is a RDNA4.5 in the wings?

There was RDNA 4.5 but that was the one that was axed

blckgrffn · Sunday at 10:42 AM

inquiss said:
There was RDNA 4.5 but that was the one that was axed

Right! Sorry.

RDNA 4.99?

inquiss · Sunday at 7:59 PM

blckgrffn said:
Right! Sorry.

RDNA 4.99?

Grey is all kinds of black and white. Tbh though if they need to cut stuff for gpgpu you could see that small incremental changes to bring forward rDNA 5 could en the way to go. Whether that's what happened or not, only adroc (supposedly) knows

Rekluse · Monday at 2:14 AM

jpiniero said:
Good catch. So def fake.

Not necessarily, AMD may still not have figured out how to nail a multi-die gaming GPU, hence no RDNA5 Halo

marees · Monday at 5:11 AM

Rekluse said:
Not necessarily, AMD may still not have figured out how to nail a multi-die gaming GPU, hence no RDNA5 Halo

They just need to get that software crossfire to work (without hitches or stutters)

leoneazzurro · Monday at 5:43 AM

marees said:
They just need to get that software crossfire to work (without hitches or stutters)

There would not be any "crossfire" - the multi-chip GPU needs to behave like a monolithic GPU software-side and that means that a significant inter-die connection bandwitdh with relatively low latency is needed.

Rekluse · Monday at 7:32 AM

The issue is the current track record

RDNA3 Launch = Borked
RDNA4 Launch = Gimped
RDNA5 Launch = ??????

Overall it's not looking promising for AMD on the chiplet GPU front. RDNA2 was good but this is a whole different ball-game with no evidence of a working concept yet

(MI300X doesn't run games remember)

igor_kavinski · Monday at 7:38 AM

Rekluse said:
(MI300X doesn't run games remember)

Any reason why it can't? Don't see why they can't scale that down to a consumer version, minus the CPU cores.

beginner99 · Monday at 8:02 AM

igor_kavinski said:
Any reason why it can't? Don't see why they can't scale that down to a consumer version, minus the CPU cores.

because latency between dies matters much less in compute. For gaming it's essential.

That and i don't even know if MI300x appears as just one GPU, earlier MI generations appears as 2 gpus to the OS which again is not very relevant for compute but very much so for gaming.

marees · Monday at 8:06 AM

Rekluse said:
Overall it's not looking promising for AMD on the chiplet GPU front. RDNA2 was good but this is a whole different ball-game with no evidence of a working concept yet

At this point I am just hoping that AMD doesn't lose the surface handheld to Intel

The Xbox next seems locked to RDNA 5
& PlayStation seems locked to Radeon for a long time
But the steam deck (this now sells half as much as xbox) could go nvidia+arm

poke01 · Monday at 8:11 AM

Rekluse said:
RDNA2 was good but this is a whole different ball-game

I really liked RDNA2, it was simple and smart architecture so much so that I got a RX 6750 XT.

This generation I planned to get RDNA3 when it got discounted. Instead, I ordered an Ada GPU taking advantage of the EOFY sales. I will post the 4070S performance in the Steel Nomad thread maybe this weekend.

I dislike the new connector on the 40 series/not providing VRAM relative to the price but another than that Ada is a great architecture compared to RDNA3 and any other GPU architecture for that matter from Apple etc.

For me it’s simple, having the best efficient architecture but also a powerful architecture will get my money. That was something AMD did with RDNA2. Ampere was good too but it was horrible due to that Samsung node efficiency wise.

It can go either way with RDNA 4, let’s see what AMD does.

igor_kavinski · Monday at 8:16 AM

beginner99 said:
That and i don't even know if MI300x appears as just one GPU

AMD Expands MI300 With GPU-Only Model, Eight-GPU Platform with 1.5TB of HBM3

AI performance to the extreme.

www.tomshardware.com

AMD conducted a demo of a 40 billion parameter Falcon-40B model running on a single MI300X GPU, but no performance metrics were provided. Instead, the LLM wrote a poem about San Francisco, the location of AMD's event. AMD says this is the first time a model this large has been run on a single GPU.

So it must be appearing to the OS as a single GPU.

leoneazzurro · Monday at 10:01 AM

igor_kavinski said:
AMD Expands MI300 With GPU-Only Model, Eight-GPU Platform with 1.5TB of HBM3

AI performance to the extreme.

www.tomshardware.com

So it must be appearing to the OS as a single GPU.

It is but as someone else said above, the amount of inter-die bandwidth and the need of a low latency connection is higher on a graphical workload (which is also less predictable than a compute workload).

adroc_thurston · Monday at 10:06 AM

beginner99 said:
That and i don't even know if MI300x appears as just one GPU

It does, but you can partition.

leoneazzurro said:
the amount of inter-die bandwidth and the need of a low latency connection is higher on a graphical workload

This is an LLM stick, latency and bandwidth are also pretty crucial there, lmao.

leoneazzurro · Monday at 10:27 AM

adroc_thurston said:
This is an LLM stick, latency and bandwidth are also pretty crucial there, lmao.

Not saying that is not important there, but IIRC you need even more in a gaming GPU due to the presence of other requirements (i.e texturing) and anyway a MI300X-like packaging should still cost a lot for consumer graphics.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Diamond Member

Lifer

Platinum Member

Diamond Member

Senior member

Junior Member

Lifer

Junior Member

Member

Diamond Member

Member

Diamond Member

Member

Member

Member

Golden Member

Member

Lifer

Diamond Member

Member

Golden Member

Lifer

Golden Member

Diamond Member

Golden Member