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

 

[Win2012R2]

It's inefficient to pay for the media engine footprint twice

Plus royalty payments to rights holders of respective commercial codes, should be fixed amount but on low end hardware that could be substantial - if even Windows was trying to ditch some of those by default, then the rest of smaller guys might also try it, only trouble in Windows it can be enabled for small fee (at least in the past that was I think).

 

[SolidQ]

RDNA4 encode support AV1 B-frame

https://twitter.com/x/status/1873018039723278858

 

[Kepler_L2]

https://videocardz.com/newz/asus-geforce-rtx-5080-and-radeon-rx-9070-xt-custom-gpu-names-leaked-16gb-memory-confirmed

9070 XT/9070 with 16GB of VRAM.

 

[jpiniero]

https://videocardz.com/newz/asus-geforce-rtx-5080-and-radeon-rx-9070-xt-custom-gpu-names-leaked-16gb-memory-confirmed

9070 XT/9070 with 16GB of VRAM.

I still say they should be only doing one model to start... but having 3 is just silly. Is the Non XT going to be only like $25 cheaper? What's the point?

 

[gaav87]

https://videocardz.com/newz/asus-geforce-rtx-5080-and-radeon-rx-9070-xt-custom-gpu-names-leaked-16gb-memory-confirmed

9070 XT/9070 with 16GB of VRAM.

Yes wasnt this known already or am i getting flashbacks ?
9070xt 256b 20gbps
9070 256b 18gbps + cut down cu's
?

 

[ToTTenTranz]

https://videocardz.com/newz/asus-geforce-rtx-5080-and-radeon-rx-9070-xt-custom-gpu-names-leaked-16gb-memory-confirmed

9070 XT/9070 with 16GB of VRAM.

Having the full 16GB on the 2nd tier sounds like excellent value... assuming it's $400 or less.

 

[Kepler_L2]

I still say they should be only doing one model to start... but having 3 is just silly. Is the Non XT going to be only like $25 cheaper? What's the point?

I think CES launch is just 9070 XT/9070, with 9060 12GB and N44 models coming later.

Yes wasnt this known already or am i getting flashbacks ?
9070xt 256b 20gbps
9070 256b 18gbps + cut down cu's
?

Yeah but not 100% confirmed.

 

[gaav87]

Having the full 16GB on the 2nd tier sounds like excellent value... assuming it's $400 or less.

Well i heard the top sku is ~749euro but may be placeholder price.

 

[ToTTenTranz]

Well i heard the top sku is ~749euro but may be placeholder price.

Yeah well, that's DOA price.


I'd even buy 2x B580 just out of spite.

 

[SolidQ]

Having the full 16GB on the 2nd tier sounds like excellent value

Only perfomance is mystery
What if that was 9070 non XT 7900GRe+ and 9070XT is ~4080?

Yeah well, that's DOA price.

We still don't know what is perfomance.

9050 vs 5060?

 

[gaav87]

Yeah well, that's DOA price.


I'd even buy 2x B580 just out of spite.

Idk i think its, a good price for highest tier aib thats with vat included.

 

[gaav87]

@adroc_thurston
More then one micro-engine or chiplet ?

 

[gaav87]

In newest amd gpuopen there is possibility of 3 ME but rdna3 and rdna4 have only 1 me would that mean 3 chiplets for rdna5 ? Any1 or am i reaching.
Or there was 3 planed for rdna4 and abandoned ?

 

[poke01]

It's inefficient to pay for the media engine footprint twice, and most AM5 CPUs have media engine (iGPU) now (the main exception is 7500F, with 8700F and 8400F in tow but those don't sell that much I think).

What about AM4??? Media engines are cheap. Especially for a $250 GPU.

Intels B580 has an excellent media engine, this is just AMD saving costs on low end.

 

[adroc_thurston]

Media engines are cheap

Yeah.

Especially for a $250 GPU.

Let's have more fun and make it $329.

Intels B580 has an excellent media engine

It also is 272mm^2. Too bad!

 

[blckgrffn]

Yeah.

Let's have more fun and make it $329.

It also is 272mm^2. Too bad!

And ironically if paired with a modern Intel CPU the B580 would be redundant with what the media engine sensitive community considers to be the best IGPU for media use available.

🤷‍♂️

 

[Jan Olšan]

What about AM4???

I know, but that platforms is kinda going out over the years these GPUs will be on the market. It's not like discounted last-gen cards won't keep being available for it, as an alternative.

Media engines are cheap. Especially for a $250 GPU.

Does this look cheap:

https://twitter.com/x/status/1848749030009937965

?

Intels B580 has an excellent media engine, this is just AMD saving costs on low end.

Yeah, but it aggravates the issue of Intel GPUs having relatively bad price performance per mm2 ratio, and that is a fundamental problem in GPUs because scaling them up involves copypasting more and more units. If you are 20-40% less dense than you need to be you have a problem that scales up with the segment of the GPU too, and in the higher tier, it caps your maximum achievable performance too.

So rationalizing die size is a prudent thing.

 

[poke01]

Does this look cheap

the media engine is the least of Intels problems with Arrow lake.. anyway Intel design is third rate now.

Yeah, but it aggravates the issue of Intel GPUs having relatively bad price per mm2 ratio, and that is a fundamental problem in GPUs because scaling them up involves copypasting more and more units. If you are 20-40% less dense than you need to be you have a problem that scales up with the segment of the GPU too, and in the higher tier, it caps your maximum achievable performance too.

So rationalizing die size is a prudent thing.

Agree.

 

[Win2012R2]

So rationalizing die size is a prudent thing.

It is prudent when you are building a top end chip, but when you are doing what AMD is doing then that's got a different name for it - cheapskating...

 

[gaav87]

So there is, a possibility gfx13 will have more than 1 GCD
More than 1 MEC and its also confirmed in amd's latest gpuopen about mes up to 3 mec's would mean up to 3 gcd interesting.

 

[Kepler_L2]

So there is, a possibility gfx13 will have more than 1 GCD
More than 1 MEC and its also confirmed in amd's latest gpuopen about mes up to 3 mec's would mean up to 3 gcd interesting.

Not GCD, they aren't doing that type of design again.

 

[gaav87]

Not GCD, they aren't doing that type of design again.

Then why would they need more than 1 mec for just 1 chip?
Hmm maybe stacking them on top of eachother idk

 

[SolidQ]

Then why would they need more than 1 mec for just 1 chip?

That Top RDNA4

 

[gaav87]

That Top RDNA4

Oh i forgot about that mlid slide so its still on the table for gfx13 or simmilar up to 3.
Well at least i have 100% confirmation 9070xt has 1mec and gfx13+ can have up to 3.
In the future when i will browse thru the kernel il keep an eye out for more than 1 me. If it has it than we talking some simmilar design as the above screen thanks.

 

[adroc_thurston]

so its still on the table for gfx13 or simmilar up to 3.

nope