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

 

[CastleBravo]

What does "+45% RT performance" even mean? It obviously can't be +45% framerate in RT or it would be stronger than a 4090. Should we interpret it as something like ~31% less performance loss when enabling RT?

 

[SolidQ]

It obviously can't be +45% framerate in RT or it would be stronger than a 4090.

it's about RE4, so it's possible, maybe RE4 love high clocks

 

[ToTTenTranz]

it's about RE4, so it's possible, maybe RE4 love high clocks

Could it be RE4 + FSR2 on the 7900XTX vs RE4 + FSR4 on the 8800XT, at "ISO quality"?

 

[SolidQ]

So 8800XT?

Loading…

videocardz.com

Could it be RE4 + FSR2 on the 7900XTX vs RE4 + FSR4 on the 8800XT, at "ISO quality"?

could be easily by diff arch and higher clocks

 

[gdansk]

Could it be RE4 + FSR2 on the 7900XTX vs RE4 + FSR4 on the 8800XT, at "ISO quality"?

It's also possibly simply made up. What's the source again? Some translated Chinese forum post by a guy with Jensen in Laker's jersey as his avatar.

 

[igor_kavinski]

It's right there on the site! https://seasonic.com/wattage-calculator/

{"cpu-modelSelect-1":{"id":"Core i9-13900KS","text":"Core i9-13900KS","tdp":"150","powerPlugs":"","recommendedPSU":0},"gpu-nameSelect-1":{"id":"Radeon RX 8800 XT","text":"Radeon RX 8800 XT","tdp":"220","powerPlugs":"2x 8-pin","recommendedPSU":550},"ssd-count":"1","fan-count":"1","hdd-count":"0","formFactor":"ATX/EPS"}

 

[Mopetar]

What does "+45% RT performance" even mean?

It means that the found one game and resolution where they could get 45% better results. I wouldn't generalize beyond that even if you accept the rumor as true.

 

[blckgrffn]

It means that the found one game and resolution where they could get 45% better results. I wouldn't generalize beyond that even if you accept the rumor as true.

8K resolution or some such nonsense possible

 

[adroc_thurston]

8K resolution or some such nonsense possible

It's a 256b part with 16G of DRAM at most. lol.

 

[branch_suggestion]

PPA looks encouraging, fingers crossed the thing is indeed the thing.

 

[Tuna-Fish]

So 8800XT?

220W sounds a little low for the top model. I suspect they'd want to juice it to the gills. If true, I wonder if this means my expectations are off, or if they are doing XTX again and XT is the rumored cut 16GB model.

 

[poke01]

People who don't care about RT and who want a decent affordable card for raster or AI compute.

What????

The 8800 XT is a better card all the way. Even for AI. No body should buy RDNA 3 unless it’s less than $300. RDNA4 is a massive leap it looks like.

 

[poke01]

Here’s a shocker, what if Ray tracing in some games is better than Nvidias last gen?

 

[CastleBravo]

220W sounds a little low for the top model. I suspect they'd want to juice it to the gills. If true, I wonder if this means my expectations are off, or if they are doing XTX again and XT is the rumored cut 16GB model.

The new leak says 25% less power than 7900 XTX for whatever N48 model roughly matches a 4080. That would be quite a bit more than 220W.

 

[igor_kavinski]

The 8800 XT is a better card all the way. Even for AI.

8800 XT won't have 24 GB RAM. Maybe 20 GB but nothing is confirmed at the moment.

If it has only 16GB, it will fail to be performant with the larger AI models.

 

[gdansk]

8800 XT won't have 24 GB RAM. Maybe 20 GB but nothing is confirmed at the moment.

If it has only 16GB, it will fail to be performant with the larger AI models.

It's 256 bit and 16GB. We have a pretty good idea of its configuration but clock rate, memory bandwidth, power limits and the impact of changes to WGP remain unknown.

 

[poke01]

8800 XT won't have 24 GB RAM. Maybe 20 GB but nothing is confirmed at the moment.

If it has only 16GB, it will fail to be performant with the larger AI models.

While the 7900 XTX does have ample VRAM for its price, RDNA3 isn’t that great of an architecture for AI. I guess it does depend on price after RDNA4 launches.

See this is the problem with not launching a 24GB RDNA4 XTX card, it would have been a great alternative to the 5080.

 

[maddie]

8800 XT won't have 24 GB RAM. Maybe 20 GB but nothing is confirmed at the moment.

If it has only 16GB, it will fail to be performant with the larger AI models.

Halo? 128GB.

 

[blckgrffn]

Halo? 128GB.

Well, you see the way they keep this 4.5ghz beast fed is with a full 512MB of infinity cache. That’s what lets it breathe through the 256bit memory interface straw.

 

[RnR_au]

If it has only 16GB, it will fail to be performant with the larger AI models.

Those that run larger AI models use multiple gpu's anyways with the model split over the gpu's. There are backends that allow you to mix and match nvidia and amd gpu's and 'smear' a single model over the cards.

 

[ToTTenTranz]

There are backends that allow you to mix and match nvidia and amd gpu's and 'smear' a single model over the cards.

Not sure how this makes much sense, as the whole thing becomes limited by the PCIe bus. And if we're just talking about having enough memory, at those token/s you might just want to get a fast 16core CPU plus lots of DDR5.

 

[RnR_au]

Not sure how this makes much sense, as the whole thing becomes limited by the PCIe bus. And if we're just talking about having enough memory, at those token/s you might just want to get a fast 16core CPU plus lots of DDR5.

If you split the model by layers, then the state of a layer that needs to be passed to the next one is tiny. Something of the order of 30KB. Its only when you split the model by 'rows' that pcie speeds may become important.

Note that I am only talking about inference, not fine tuning or training. For these latter tasks, pcie speeds or even nvlink if you have it becomes important.

Edit: a source where this is talked about - https://github.com/turboderp/exllama/discussions/16#discussioncomment-6245573

 

[GodisanAtheist]

Why are we all acting like 5700XT part deux is going to be some sort of dark horse AI champ. Did I miss something?

 

[HurleyBird]

It's a 256b part with 16G of DRAM at most. lol.

Wouldn't be surprised by a 32GB "Budget AI, but we'll mark it up quite a bit higher than the cost of the memory" SKU at some point, if it's possible.

 

[RnR_au]

Wouldn't be surprised by a 32GB "Budget AI, but we'll mark it up quite a bit higher than the cost of the memory" SKU at some point, if it's possible.

It would be awesome if such a thing was released, but folk running AI models at home is a tiny niche market, but given the size of the homelab hobbyists group maybe someone will take a punt.

Edit: what some local AI folk are currently buying in the 32GB vram class - https://www.ebay.com.au/itm/125006475381