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

 

[SolidQ]

https://twitter.com/x/status/1876624801600532711

 

[adroc_thurston]

AMD hasn’t been nailing the execution of anything after Zen 3 and RDNA2

It's nice to live in the world without Genoa or Turin.
Granite Rapids is the comeback of the century right?

 

[gdansk]

Just lower the price of the current generation

They are not repeating the Navi 2 inventory problem this time. For example, the 7900 GRE is already gone in the US.

 

[insertcarehere]

9070 non XT is 16GB, Seems 9600XT is N48 12gb version

For all our sakes that very chonky cooler and 3x 8-pin better be ludicrous overkill rather than indicative of actual power draw. Pretty sure all or nearly all cards based on N32 only had 2x 8-pin setups.

 

[JustViewing]

If AMD really wants to sell more graphics card, they can simply increase the memory. If 9070XT have like 48GB it will sell like hot cakes because of AI craze. This would in turn in help AMD in open source AI frameworks. AMD even boasting about Zen 5 Halo CPU is 2.4 times faster than NVidia 4090 in 70GB models.

 

[gdansk]

If 9070XT have like 48GB it will sell like hot cakes because of AI craze

I don't think it'll work. Intel might be trying this though, so let's see.

 

[adroc_thurston]

This would in turn in help AMD in open source AI frameworks

They already work with Meta on that. Next.

 

[SolidQ]

For all our sakes that very chonky cooler and 3x 8-pin better be ludicrous overkill rather than indicative of actual power draw. Pretty sure all or nearly all cards based on N32 only had 2x 8-pin setups

9070xt is 3x8, non xt is 2x8

 

[JustViewing]

They already work with Meta on that. Next.

How that is related to higher RAM and helping open source?

 

[adroc_thurston]

How that is related to higher RAM and helping open source?

Because Pytorch is already open-source and getting higher ram than MI325X is hard.

 

[Meteor Late]

If AMD really wants to sell more graphics card, they can simply increase the memory. If 9070XT have like 48GB it will sell like hot cakes because of AI craze. This would in turn in help AMD in open source AI frameworks. AMD even boasting about Zen 5 Halo CPU is 2.4 times faster than NVidia 4090 in 70GB models.

And that's exactly why they won't do it, because they want to sell you Strix Halo for that exact purpose.

 

[gdansk]

And that's exactly why they won't do it, because they want to sell you Strix Halo for that exact purpose.

If AMD had their way you would rent a MI300 of some sort instead.

 

[eek2121]

If AMD had their way you would rent a MI300 of some sort instead.

If AMD big tech companyhad their way you would rent a MI300 of some sort their product instead.

Fixed that for ya! 🤣

 

[gdansk]

If AMD big tech companyhad their way you would rent a MI300 of some sort their product instead.

Fixed that for ya! 🤣

But especially so in this case. They saved money focusing software support for ML/HPC on CDNA products.

 

[JustViewing]

Because Pytorch is already open-source and getting higher ram than MI325X is hard.

This has no relation to what I mentioned. How, having 48GB in consumer card any connection to Pytorch and MI325X super computer/server part? Having 48GB in AMD consumer card will bring loads of new developers. You know, most of the open source AI tools/framework doesn't work optimally or not at all with AMD graphics card.

 

[adroc_thurston]

How, having 48GB in consumer card any connection to Pytorch and MI325X super computer/server part?

Because that's the relevant open-source framework for ML.

You know, most of the open source AI tools/framework doesn't work optimally or not at all with AMD graphics card.

Consumer ones indeed

 

[JustViewing]

Because that's the relevant open-source framework for ML.

Consumer ones indeed

Your replies doesn't seems to make any connection to what I posted. Anyway, I'll stop here.

 

[gaav87]

let’s phase it another way, it can work with RDNA3 but AMD chose RDNA4 as the base.

I’m not against that, just as long as all RDNA4 cards support FSR4

No rdna3 has no FP8. It can work on FP16 but 4x slower or 8x if using swmma

 

[gaav87]

If AMD really wants to sell more graphics card, they can simply increase the memory. If 9070XT have like 48GB it will sell like hot cakes because of AI craze. This would in turn in help AMD in open source AI frameworks. AMD even boasting about Zen 5 Halo CPU is 2.4 times faster than NVidia 4090 in 70GB models.

AI exp is so bad on AMD i gave up like 8months ago even bing is better... In the begining of AI craze like 2+ years ago u litterary had to know python to use it with way way way way worse results and perf vs rtx 3000 and 4000. You have so many models for nvidia its not even fair.

 

[JustViewing]

AI exp is so bad on AMD i gave up like 8months ago even bing is better... In the begining of AI craze like 2+ years ago u litterary had to know python to use it with way way way way worse results and perf vs rtx 3000 and 4000. You have so many models for nvidia its not even fair.

That is exactly the point I was making, if there is a consumer card with huge amount of memory it will encourage more AI developers/researchers to buy AMD cards. This in turn will benefit AMD and consumers as well.
BTW, Amuse and LMStudio works flawlessly with Radeon graphics cards with single click installation. ComfyUI with Zluda also works most of the time as it is.

 

[Heartbreaker]

That is exactly the point I was making, if there is a consumer card with huge amount of memory it will encourage more AI developers/researchers to buy AMD cards. This in turn will benefit AMD and consumers as well.
BTW, Amuse and LMStudio works flawlessly with Radeon graphics cards with single click installation. ComfyUI with Zluda also works most of the time as it is.

If memory for AI is the demand, that is what Strix Halo is all about. They covered, this at the press event, where it seemed primarily aimed at AI. With up to 128 GB of RAM, 96GB useable by the GPU for AI. This blows away any GPU.

 

[gdansk]

That is exactly the point I was making, if there is a consumer card with huge amount of memory it will encourage more AI developers/researchers to buy AMD cards. This in turn will benefit AMD and consumers as well.

I'm not following, why would they do that? Because their model doesn't fit in 32GB? I know some people like a challenge but I don't think it's significant enough to justify switching to clamshell 24gb GDDR6 chips that don't seem to exist.

 

[JustViewing]

I'm not following, why would they do that? Because their model doesn't fit in 32GB? I know some people like a challenge but I don't think it's significant enough to justify switching to clamshell 24gb GDDR6 chips that don't seem to exist.

Well, as I mentioned in the previous post if AMD wants to increase the sale of 9070 they can do that by increasing the amount RAM. Whether it is practical or not, I have no idea. But it would benefit AMD in the long run as more AI devs and students will buy this card (of course if it cost competitive only).

If memory for AI is the demand, that is what Strix Halo is all about. They covered, this at the press event, where it seemed primarily aimed at AI. With up to 128 GB of RAM, 96GB useable by the GPU for AI. This blows away any GPU.

That is true, however 9070 is a more powerful card than Strix Halo. I think for AI multi-card solution works. So 2 48GB cards can function as 96GB card. Anyway, I doubt AMD will do this. They may release a Pro version later with higher markup.

 

[Win2012R2]

So 2 48GB cards can function as 96GB card

You need fast interconnect between them to make it work like one 96 GB card, otherwise it's 2 48 GB cards

With up to 128 GB of RAM, 96GB useable by the GPU for AI. This blows away any GPU.

It's not very fast memory

 

[Bryo4321]

BTW, Amuse and LMStudio works flawlessly with Radeon graphics cards with single click installation. ComfyUI with Zluda also works most of the time as it is.

yeah Amuse is awesome. Works about as well for me on a 7900xtx as SDGUI did when I had a 4080. Love the 1 click model downloads