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

 

[itsmydamnation]

7900XTX sold out, all of them in Australia due to failure of RTX 50

I'll be getting the 9070XT day 1

i got mine 4 months ago for 1300AUD ..... feeling pretty good about now 🙂

 

[poke01]

i got mine 4 months ago for 1300AUD ..... feeling pretty good about now 🙂

Thats a very good price. You're set for a long while esp with the 24GB VRAM

 

[gaav87]

GDS と GWS を持たない AMD RDNA 4 GPU | Coelacanth's Dream

AMD GCN アーキテクチャから存在していた、すべての CU (Compute Unit) からアクセス可能なスクラッチメモリ GDS (Global Data Share) と、異なる CU や SE (Shader Engine) で実行されているスレッドグループ、Wave と同期

www.coelacanth-dream.com

Is this accurate they got rid of GDS and GWS ? So last terascale GCN remnants from rdna4 ?
Wouldnt this mean higher software overhead more complex scheduler. Guess the added new GL1 would mitigate shared memory and Inter-wavefront Communication.
(I have 38C fever so i may be talking bs)

 

[Kepler_L2]

Is this accurate they got rid of GDS and GWS ?

Old news and doesn't really matter for any modern shader workload.

 

[SolidQ]

if (info->gfx_level >= GFX10 && info->gfx_level < GFX12) fprintf(f, " l1_cache_size = %i KB\n", DIV_ROUND_UP(info->l1_cache_size, 1024));

If i understand right 1mb L1 cache size?

 

[Keller_TT]

if (info->gfx_level >= GFX10 && info->gfx_level < GFX12) fprintf(f, " l1_cache_size = %i KB\n", DIV_ROUND_UP(info->l1_cache_size, 1024));

If i understand right 1mb L1 cache size?

That piece of code just outputs L1 cache size if the arch is GFX10 or 11. It's converting bytes into KB by dividing it by 1024 and will display l1_cache_size = some KB. Looks like data log into a file stream.

 

[gaav87]

if (info->gfx_level >= GFX10 && info->gfx_level < GFX12) fprintf(f, " l1_cache_size = %i KB\n", DIV_ROUND_UP(info->l1_cache_size, 1024));

If i understand right 1mb L1 cache size?

No gl1 128->256kb if i remember right

 

[SolidQ]

No gl1 128->256kb if i remember right

Strange, TPU specs show 256 in RDNA3

NV helping sell Radeon?

RX 7900 XTX is rapidly going out of stock across major retailers following the RTX 50 series launch

Following the RTX 50 series launch, it seems as if many have started opting for AMD's 7900 XTX instead, leading to stock-outs for the card.

www.pcguide.com

 

[jpiniero]

Strange, TPU specs show 256 in RDNA3

NV helping sell Radeon?

Newegg is basically bone dry of nVidia GPUs above 4060 Ti. And it's been this way for awhile.

 

[marees]

Valve adds RDNA 4 support to Steam OS

A Valve Steam console with an AMD RX 9070 GPU is rumored to be in the works, but it looks unlikely

Speculation about Valve's SteamOS follow up to the Steam Deck has taken off online, but it looks more like a case of extreme enthusiasm.

www.pcgamesn.com

 

[gaav87]

Strange, TPU specs show 256 in RDNA3

NV helping sell Radeon?

RX 7900 XTX is rapidly going out of stock across major retailers following the RTX 50 series launch

Following the RTX 50 series launch, it seems as if many have started opting for AMD's 7900 XTX instead, leading to stock-outs for the card.

www.pcguide.com

Idk im super sick but i guess the new gl cache that is between 1 and 2 is either 128 or 256. So they got rid of 64 GDS and the new cache is taking its place ?

 

[adroc_thurston]

doesn't really matter for any modern shader workload

But will it break New Vegas? Only one way to find out!

Wouldnt this mean higher software overhead more complex scheduler

No it just means some things have to join the dodo.
CDNA axed these a whole ago.

So they got rid of 64 GDS and the new cache is taking its place ?

No, GDS is gone. All the _global stuff went to GMEM thru L2 forever anyway.

 

[poke01]

What are peoples estimates on the Blender RT score using zluda for RDNA4?

I think it will be better than 7900 XTX

 

[poke01]

https://videocardz.com/newz/amd-ceo-confirms-radeon-9070-will-go-on-sale-early-march

Hopefully lots of supply, at least be better than Nvidia

 

[marees]

https://videocardz.com/newz/amd-ceo-confirms-radeon-9070-will-go-on-sale-early-march

Hopefully lots of supply, at least be better than Nvidia

Early March Launch for the 9070 series

 

[soresu]

What are peoples estimates on the Blender RT score using zluda for RDNA4?

I think it will be better than 7900 XTX

I more interested in future improvement of HIP RT (or perhaps more cross platform Intel alternatives sh.

Zluda is at best only going to translate the non HW RT accelerated backend, which is only going to see less and less development as time goes on vs the main Optix backend.

 

[Ghostsonplanets]

Don't be so confident. I think AMD had originally intended this as a small launch.

They do have some extra time to prepare given the delay.

Valve adds RDNA 4 support to Steam OS

A Valve Steam console with an AMD RX 9070 GPU is rumored to be in the works, but it looks unlikely

Speculation about Valve's SteamOS follow up to the Steam Deck has taken off online, but it looks more like a case of extreme enthusiasm.

www.pcgamesn.com

People are taking this as Valve doing a SteamConsole regen. But isn't this just support for future SteamOS Desktop release?

https://videocardz.com/newz/amd-ceo-confirms-radeon-9070-will-go-on-sale-early-march

Hopefully lots of supply, at least be better than Nvidia

Good. They will take advantage of Nvidia paper launch period.

Honestly, Navi 44 should have been ready too. AMD would basically offer top to bottom SKUs while Nvidia supply is small.

 

[Kepler_L2]

People are taking this as Valve doing a SteamConsole regen. But isn't this just support for future SteamOS Desktop release?

It's just Valve supporting open source Linux Vulkan drivers like usual.

 

[ToTTenTranz]

It's just Valve supporting open source Linux Vulkan drivers like usual.

Fremont is still a thing, though it might be just a small set-top box with the Van Gogh chip.

 

[marees]

Fremont is still a thing, though it might be just a small set-top box with the Van Gogh chip.

That test device had a discrete mobile GPU equivalent to a cut down strix halo. I guess it depends on if valve gets the better deal on 6600s or the cut down strix halo

 

[Kepler_L2]

Fremont is still a thing, though it might be just a small set-top box with the Van Gogh chip.

That was using Lilac board, which is the Rembrandt socket. Valve was likely testing how the Lenovo Legion Go S performed with Proton.

 

[branch_suggestion]

GL1キャッシュがバッファへと変わる AMD RDNA 4 アーキテクチャ | Coelacanth's Dream

読み取り専用キャッシュとして RDNA 1 アーキテクチャから導入され、RDNA 3 アーキテクチャからは 256KiB (RDNA 1/2 は 128KiB) に増量された GL1キャッシュ (Graphics L1キャッシュ) だが、RDNA 4 からはキャッシュではなくバッファへと変わる。

www.coelacanth-dream.com

Nobody has posted this yet, anywhere.

Now I'm starting to remember some interesting research from years ago.
If elements of this shared L1 paradigm are in RDNA4 that would be very cool indeed.

https://adwaitjog.github.io/docs/pdf/sharedl1-pact20.pdf

 

[ToTTenTranz]

That test device had a discrete mobile GPU equivalent to a cut down strix halo. I guess it depends on if valve gets the better deal on 6600s or the cut down strix halo

Lilac is just a test motherboard from AMD, as mentioned above. Some confusion appeared when Fremont appeared with a Lilac board because some there are some geekbench results with a 6600S GPU, but others without.

That was using Lilac board, which is the Rembrandt socket. Valve was likely testing how the Lenovo Legion Go S performed with Proton.

Valve's logs show them getting development input from Quantum Computer which is the Steam Deck's manufacturer. Is Quantum Computer making the Legion Go S?



There's also mention of implementing HDMI CEC in Fremont, which the Legion Go S doesn't have.

 

[marees]

Frank Azor says AMD might relase full details of RDNA 4

https://twitter.com/x/status/1887182628308029507

 

[Bigos]

"soon" -> "in a month"?