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

 

[igor_kavinski]

The API to allow for a non-serial approach doesn't exist yet

Doesn't DirectX 12 have special support for multiple GPUs? How hard could it be, by virtualizing at the driver level, a single graphics card into multiple GPUs matching the number of GPU chiplets on the card?

 

[adroc_thurston]

Doesn't DirectX 12 have special support for multiple GPUs?

Multiadapter is dead.

How hard could it be, by virtualizing at the driver level, a single graphics card into multiple GPUs matching the number of GPU chiplets on the card?

Yay we revived crossfire!
genius.

 

[igor_kavinski]

Yay we revived crossfire!
genius.

Yes but it's 4-way Crossfire between the GPU chiplets ON a single PCIe 5.0 graphics card

It can even be one big GPU chiplet in the middle and four smaller chiplets to do foveated rendering, as the human eye isn't going to be focusing on the entire screen all at once so the GPU processing can be concentrated on where the eyes are looking at. So that can be a region around the third person or around the crosshairs in a first person shooter.

 

[adroc_thurston]

Yes but it's 4-way Crossfire between the GPU chiplets ON a single PCIe 5.0 graphics card

Won't work.

It can even be one big GPU chiplet in the middle and four smaller chiplets to do foveated rendering, as the human eye isn't going to be focusing on the entire screen all at once so the GPU processing can be concentrated on where the eyes are looking at. So that can be a region around the third person or around the crosshairs in a first person shooter.

Esoteric stuff that will never work.
Only the basic most straightforward approach works.

 

[igor_kavinski]

Esoteric stuff that will never work.

Speaking from experience?

Only the basic most straightforward approach works.

Then what's the way forward?

 

[adroc_thurston]

Speaking from experience?

Yea.

Then what's the way forward?

the RDNA4 way, duh.

 

[igor_kavinski]

Yea.

More details needed please. We enjoy stories from inside the trenches

the RDNA4 way, duh.

Monolithic die? Or are you referring to something else?

 

[branch_suggestion]

I imagine the simplest way is disabling all but 1 AID when running legacy APIs.

 

[adroc_thurston]

I imagine the simplest way is disabling all but 1 AID when running legacy APIs.

Every single graphics API ever made is a legacy one.

 

[PJVol]

Besides the rumours about 8000 gen showcase at CES, what's the current state of affair on what the RDNA4 lineup is in essence?
So far I see only two "leakers" presented their view of it:
* MLID saying there are midrange n43 and n44 only, and somehow n43 is gonna perform between 7900xt and 4080 super. ?????
* RGT said there are n48, n44 - midrange (8700 & 8600), and recently, some mysterious n31 refresh for the top part 8800 with the 7900gre like performance (???) and n48 (aka mlid's n43) is 32WGP (so no way it could have the 7900XT perf. as per mlid). Also he didn't mention if this n31 refresh is monolithic or not.

WTH? What 4080 super they are talking about? Anyone got an idea what's going on in the Radeon Tech labs and/or headquaters?

 

[jpiniero]

Besides the rumours about 8000 gen showcase at CES, what's the current state of affair on what the RDNA4 lineup is in essence?
So far I see only two "leakers" presented their view of it:
* MLID saying there are midrange n43 and n44 only, and somehow n43 is gonna perform between 7900xt and 4080 super. ?????
* RGT said there are n48, n44 - midrange (8700 & 8600), and recently, some mysterious n31 refresh for the top part 8800 with the 7900gre like performance (???) and n48 (aka mlid's n43) is 32WGP (so no way it could have the 7900XT perf. as per mlid). Also he didn't mention if this n31 refresh is monolithic or not.

Best to just ignore those fake noodz guyz.

 

[igor_kavinski]

Every single graphics API ever made is a legacy one.

You can't call all of them legacy unless there's one on the horizon that will best them all. Which is it? Some new fangled API brewing in secrecy? RDNA4 will support that?

 

[Tup3x]

I can see chiplet approach working only if developers can treat such a device the same as monolithic.

 

[igor_kavinski]

I can see chiplet approach working only if developers can treat such a device the same as monolithic.

Or the drivers are so good that they make the chiplets work in tandem and appear monolithic to game engines.

 

[maddie]

Too bad cooling problems prevents going vertical with SOIC. Would solve the (excessive data movement / energy cost) problems.

 

[PJVol]

Best to just ignore those fake noodz guyz.

So, basically no one knows anything about RDNA4, other than some mysterious navi4c has been canceled?

Yeah, unfortunately those who silently knows and aren't used to talk a lot remain silent, unlike so-called YT "leakers", whom they like to catch in lies or mispredictions, however, after the product launch.

 

[Ajay]

So, basically no one knows anything about RDNA4, other than some mysterious navi4c has been canceled?

Yeah, unfortunately those who silently knows and aren't used to talk a lot remain silent, unlike so-called YT "leakers", whom they like to catch in lies or mispredictions, however, after the product launch.

Generally, the more one is in the know, the less they talk. MILD and RTG throw rumor after rumor up against a wall till something sticks (like spaghetti). The they ignore all the bogus stuff they said in the past and point out that they got it right. Some of the more reliable twitterati are Kepler_L2, kopite7kimi and a few others depending on the product segment. Pretty much all the major tech firms have worked to stifle leaks and give us less detailed roadmaps. Such is life.

 

[GodisanAtheist]

Anything ever come of that one AMD guy that claimed that high end RDNA had not been canceled?

 

[Ajay]

Anything ever come of that one AMD guy that claimed that high end RDNA had not been canceled?

Nah. One post and gone. He was an employee of AMD RTG. His bosses probably found out and had a bit of a talk with him. So, we'll just have to see what comes out and decide whether it 'high performance' or not.

 

[jpiniero]

Anything ever come of that one AMD guy that claimed that high end RDNA had not been canceled?

An N31 refresh could be considered high end.

 

[Ajay]

An N31 refresh could be considered high end.

Well, it seems that the work needed to fix N31 is far more than just a 're-spin'. I just doubt that AMD would spend the time, money and engineering talent on such a project when they have some much else going on in RTG. It seems, as with RDNA4, that they are going to just keep rolling forward and not let missed targets be a cause for pulling back on next gen products. We'll just have to see how well this works out for them (financially).

 

[adroc_thurston]

You can't call all of them legacy unless there's one on the horizon that will best them all.

Well yea, CUDA/ROCm beat the snot out of graphics APIs.
Those are not legacy.

Some new fangled API brewing in secrecy? RDNA4 will support that?

Graphics is inherently legacy, can't do anything about it.

An N31 refresh could be considered high end.

Not happening (at least not anymore).

 

[igor_kavinski]

Graphics is inherently legacy, can't do anything about it.

So legacy APIs are here to stay, right? What is RDNA4 doing better in that department? Is it somehow doing graphics through a wrapper over ROCm? As in 3D graphics pipeline made to run as compute?

 

[adroc_thurston]

So legacy APIs are here to stay, right?

Yea, forever.

What is RDNA4 doing better in that department?

It's a GPU that does GPU things?

Is it somehow doing graphics through a wrapper over ROCm?

Maybe in like 10 years.

 

[DisEnchantment]

Hmmm GFX12 are showing up.

[AMDGPU] Define new targets gfx1200 and gfx1201 by jayfoad · Pull Request #73133 · llvm/llvm-project

Define target names and ELF numbers for new GFX12 targets gfx1200 and gfx1201. For now they behave identically to GFX11.

github.com

Nothing special yet.