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

 

[GodisanAtheist]

That's how R&D cycles go.
You have the problem and you make a solution.

-Do you make a product and stack before you have a solution?

 

[Mopetar]

-Do you make a product and stack before you have a solution?

If you think you have one you might just do that. Even if it's tricky, the financial benefits of AMD pulling it off may make the risk seem reasonable.

It makes more sense for CDNA to go down this line as distributed loads don't need a monolithic die any more than they need a single die. For consumer GPUs it remains a problem that may be too difficult solve.

 

[adroc_thurston]

Do you make a product and stack before you have a solution?

The product is the solution.

 

[jpiniero]

The product is the solution.

I think the question is why AMD would go down this path without feeling like they have solved the problems before committing. Unless AMD simply doesn't care about the high end and is OK if they cancel products if they can't get it working in time.

 

[PJVol]

Unless AMD simply doesn't care about the high end and is OK if they cancel products if they can't get it working in time.

Just curious, how much could navi4c mask (or masks) cost?

 

[maddie]

I think the question is why AMD would go down this path without feeling like they have solved the problems before committing. Unless AMD simply doesn't care about the high end and is OK if they cancel products if they can't get it working in time.

Maybe because when you commit to a path, you're working on assumptions of a solution being found based on the present data. This happens all the time in many fields of advanced engineering. Hi-tech is always a balance on what will be thought to have been solved by roll-out.

 

[linkgoron]

I think the question is why AMD would go down this path without feeling like they have solved the problems before committing. Unless AMD simply doesn't care about the high end and is OK if they cancel products if they can't get it working in time.

It's not like this is the first time AMD has committed to something that didn't really work. HBM1 and Fiji are such an example.

 

[adroc_thurston]

HBM1 and Fiji are such an example.

HBM is literally in every HPC part out there.

I think the question is why AMD would go down this path without feeling like they have solved the problems before committing

Semis R&D cycles are always chains of rather bold bets and this was one of them.
Again, N4C was canned for TTM reasons and not anything technical.

 

[linkgoron]

HBM is literally in every HPC part out there.

yes, now. Fiji was a failure. AMD overcommitted to a costly, difficult technology before it was ready for prime time. I don't think we've seen any consumer card with HBM since?

Nvidia completely destroyed them with the 9 series, and the following 10 series. I also believe that they sell more cards with HBM than AMD, today.

 

[jpiniero]

Semis R&D cycles are always chains of rather bold bets and this was one of them.
Again, N4C was canned for TTM reasons and not anything technical.

Is there any reason to think they've been able to solve the issues with scalable chiplets though.

 

[SolidQ]

That really strange. We can beat 7900XT, but not sure if can beat 4080 super 😁

Spoiler

 

[branch_suggestion]

So we return bac to my OGP about topic. It is cost and revenue opportunity. Everything is finite resource, hence use that limited resource as efficiently as you can - produce MI300 and MI400 as much and as fast as you can. There is AI craze going on out there.

RDNA and CDNA use largely different packaging streams, that is the point.
NaviC uses a regular substrate for the base dies, everything above the base dies uses advanced packaging, compared to HBM parts which require big interposers above the substrate. This is cheaper, yields better and ultimately allows for higher volume. This will extend across every product division eventually.

 

[Timorous]

I have said it before.

I can get that 4c was a thing and was canned due to ttm. I get it was a win big style part so would have had a 4090 or higher asking price.

What I don't get is how that means the next highest part in the stack is no better than 7900XT/XTX tier. That bit makes zero sense to me

A stack with 4c and super complex packaging at the top with an RDNA3 esque MCM config sitting below that super halo part and then some monolithic parts for low end and laptops makes sense. The huge gap, not so much.

 

[adroc_thurston]

Fiji was a failure. AMD overcommitted to a costly, difficult technology before it was ready for prime time

Oh there's a lot more to that part of Radeon history but not public stuff.

Is there any reason to think they've been able to solve the issues with scalable chiplets though.

Yea it's called MI300.
There's only one real issue with N4C and that's validation that may be protracted enough to put the end result close enough to RDNA5 parts to not matter.

 

[branch_suggestion]

I have said it before.

I can get that 4c was a thing and was canned due to ttm. I get it was a win big style part so would have had a 4090 or higher asking price.

What I don't get is how that means the next highest part in the stack is no better than 7900XT/XTX tier. That bit makes zero sense to me

A stack with 4c and super complex packaging at the top with an RDNA3 esque MCM config sitting below that super halo part and then some monolithic parts for low end and laptops makes sense. The huge gap, not so much.

There were 5 parts, 4C/X/M and the 2 mono parts that made the cut.
4X was probably intended to be faster than N31, 4M might've been similar. AMD was going for a true full stack solution, now we will need to wait until RDNA5.

 

[jpiniero]

Yea it's called MI300.
There's only one real issue with N4C and that's validation that may be protracted enough to put the end result close enough to RDNA5 parts to not matter.

Compute is one thing, making it work with games is another.

 

[linkgoron]

Oh there's a lot more to that part of Radeon history but not public stuff.

Is this just bragging that you know stuff? Either say what you know or don't say anything. You're not adding anything to the conversation.

The fact is, AMD over-committed to HBM at the time. It was the only "real" consumer card that ever used the memory (unless we consider the $3000 Titan V, which I believe we can ignore). Nvidia, on the other hand, waited for the right time to use HBM and used it in more fitting products. HBM on the Fury cards was too small to be competitive in an era where Nvidia high-end cards had 8GB memory (as well as AMD's own 390 cards), and Fiji itself was slower than the 980TI which was released first, and was an overclocking monster.

Don't see why you had to comment against my example, HBM on Fiji was a premature and dumb decision by AMD. It took them years to create a new architecture that actually managed to outperform Fiji, and they only managed to become competitive again with RDNA and RDNA2. HBM panned out, but mostly for Nvidia, not AMD (Yes, AMD and Intel also use it today).

 

[jpiniero]

Don't see why you had to comment against my example, HBM on Fiji was a premature and dumb decision by AMD.

It's great at mining, isn't it?

 

[adroc_thurston]

Compute is one thing, making it work with games is another.

Eh the difficult uArch decisions are roughly the same for both; just less requirements for the specific way of doing stuff correctly for CDNA.

AMD over-committed to HBM at the time

They've actually undercommited.

 

[DaaQ]

Them Vega cards also used HBM, 16GB if memory serves. What was it Frontier editions?

 

[jpiniero]

Eh the difficult uArch decisions are roughly the same for both; just less requirements for the specific way of doing stuff correctly for CDNA.

Well, they are gaming products. If it doesn't work right with games... that's not going to fly.

 

[adroc_thurston]

Well, they are gaming products. If it doesn't work right with games... that's not going to fly.

It's not about if, but when.

 

[Ajay]

Yeah, but that's not an MS issue, it's a legacy designs from ca 1995 issue.
OGL and derivatives are the same.

Fair enough. Legacy stuff always keeps getting dragged along for decades - the golden handcuffs of backward compatibility.

 

[igor_kavinski]

Legacy stuff always keeps getting dragged along for decades - the golden handcuffs of backward compatibility.

What I don't get is, why they don't just create a wrapper/compatibility layer for the legacy stuff and move on? The older stuff works at 100+ fps anyway. A little overhead of even 50 fps slowdown isn't gonna make gamers come out with pitchforks. Who in their right mind wants to buy the latest $1000 GPU to add another 100 fps to their 300 fps legacy game?

 

[moinmoin]

What I don't get is, why they don't just create a wrapper/compatibility layer for the legacy stuff and move on? The older stuff works at 100+ fps anyway. A little overhead of even 50 fps slowdown isn't gonna make gamers come out with pitchforks. Who in their right mind wants to buy the latest $1000 GPU to add another 100 fps to their 300 fps legacy game?

Everything's serial right now so everything would be legacy. The API to allow for a non-serial approach doesn't exist yet, so games can't even start to move over yet. And once that exists that will still take time, and all already launched games are very unlikely to ever make that move.