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]

Ok Kepler post RDNA4 RT

https://twitter.com/x/status/1814966230005985401

Double Intersect Engine it's like this?

4/1 to 8/2?

 

[ryanjagtap]

Ok Kepler post RDNA4 RT

https://twitter.com/x/status/1814966230005985401

Double Intersect Engine it's like this?

4/1 to 8/2?

Couldn't they offload the BVH traversal to the AI accelerators introduced in RDNA3 or does the PS5 Pro die does not include AI accelerators? So due to cross compatibility issues it will still run on the shaders?

 

[Saylick]

Ok Kepler post RDNA4 RT

https://twitter.com/x/status/1814966230005985401

Double Intersect Engine it's like this?

4/1 to 8/2?

Yep.

https://twitter.com/x/status/1769514652239589885

 

[SolidQ]

Yep.

Interesting how blackwell going 8/4?

 

[soresu]

What's with the redacted stuff in Kepler's tweet?

 

[Ghostsonplanets]

What's with the redacted stuff in Kepler's tweet?

RDNA 4 non-RT new features/improvements.

Including > 3.x GHz nominal clocks

 

[soresu]

Couldn't they offload the BVH traversal to the AI accelerators introduced in RDNA3 or does the PS5 Pro die does not include AI accelerators? So due to cross compatibility issues it will still run on the shaders?

GPUOpen published a paper on something like that as a software technique, though for intersection rather than traversal:

https://gpuopen.com/download/publications/HPG2023_NeuralIntersectionFunction.pdf

 

[soresu]

RDNA 4 non-RT new features/improvements.

Including > 3.x GHz nominal clocks

8 features besides that if I count correctly.

 

[Saylick]

RDNA 4 non-RT new features/improvements.

Including > 3.x GHz nominal clocks

Looks like 3.4 GHz, which would be sweet.

https://twitter.com/x/status/1814983542079676852

 

[Ghostsonplanets]

Looks like 3.4 GHz, which would be sweet.

https://twitter.com/x/status/1814983542079676852

3.4GHz (I assume for Navi 48) would be pretty crazy. I assume the target is 4+GHz for RDNA 5 generation.

 

[Saylick]

3.4GHz (I assume for Navi 48) would be pretty crazy. I assume the target is 4+GHz for RDNA 5 generation.

Let’s not get too crazy… *hears the whistle of the RDNA 5 hype train about to depart*

 

[gdansk]

I've been on the 3500MHz hype train for months.

And I'm not getting off until after it collides with reality.

 

[soresu]

And I'm not getting off until after it collides with reality.

All aboard wid ya!

 

[poke01]

I’m not boarding the train this time, I pre-booked for RDNA 5 and Blackwell-next tho.

 

[soresu]

and Blackwell-next tho

Pretty sure that's been named now, starts with an R.

 

[soresu]

Ah it's Rubin, after Vera Rubin.

I think that they could have found someone a bit more impressive tbh like Marie Curie.

 

[Hans Gruber]

Nobody is boarding the RDNA4 train. AMD needs to get rid of the maintaining margins thinking. Revenue is more important than profits if your desire for profit destroys your brand with inferior products. AMD drivers have always been passably bad. They always fix them, break them and fix them again. They have a 3 steps forward two steps back approach to drivers. Meanwhile Nvidia occasionally releases a bad driver but they release new drivers every week vs. every month from AMD.

The glory days of AMD in recent memory were RDNA1. It made Nvidia users think about value vs. cost. If you provide customers a good consistent value. They are more likely to buy from you again and upgrade more often.

If people think I am wrong. Look at the most recent steam survey. The middle of the road GPU's are what most steam users have. AMD has been around a long time. There is no excuse for them not offering a 5080 competitor with RDNA4.

 

[poke01]

The glory days of AMD in recent memory were RDNA1.

RDNA1 was crap, you mean RDNA2?

 

[Hans Gruber]

RDNA1 was crap, you mean RDNA2?

I am talking price/performance only. In RDNA2 it would be the 6800xt. The argument is that AMD cards do one thing well and only one thing, gaming. Everything else is crap. Hopefully that changes starting with RDNA4 and beyond.

The RX 5700XT is still well loved in the value gaming community. It's old now but offered a lot of performance for the price.

 

[DisEnchantment]

What is the consensus on the process node for RDNA4?
Would it be N4P or N3E?

Lots of patents around those items listed by Kepler from AMD

 

[Aapje]

Ah it's Rubin, after Vera Rubin.

I think that they could have found someone a bit more impressive tbh like Marie Curie.

For the people who pick these names, achievement is one of the less important elements. I probably can't discuss which elements matter more.

 

[Aapje]

What is the consensus on the process node for RDNA4?
Would it be N4P or N3E?

N4P, probably. The companies seem to act like N3 doesn't exist, so it's probably no good. If Nvidia doesn't use it, then AMD certainly won't either.

 

[soresu]

For the people who pick these names, achievement is one of the less important elements. I probably can't discuss which elements matter more.

Not saying Rubin has no achievements at all - but IMHO her most notable being evidence for the existence of dark matter, but not identification of, or proof of it.

To me unless dark matter gets confirmed tomorrow it amounts to little more than the scientific equivalent of a meme, much as I feel about string theory and related hypothetical physics.

By comparison Marie Curie's achievements are much more tangible and historic.

Plus I mean, the sacrifice of health she gave (knowingly or otherwise) to achieve them is no small thing on top of that.

 

[ryanjagtap]

N4P, probably. The companies seem to act like N3 doesn't exist, so it's probably no good. If Nvidia doesn't use it, then AMD certainly won't either.

N4P for RDNA4 seems plausible.
RDNA4 is a mid range competing product lineup. It would need volume and N4P will be cheaper and still give some reduction to die size compared to the MCM RDNA3 counterparts.
N3E is used by AMD, but it's all going to Instinct GPUs and Turin-Dense. More $/mm2 in those products.

 

[poke01]

The companies seem to act like N3 doesn't exist, so it's probably no good

N3E is good. Qualcomm, Apple, AMD are using it this year. Its just costly