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

 

[Abwx]

From what I read it means clock for clock the performance increase is 11% N5 vs. N4P and a 22% reduction in power/efficiency N5 vs N4P.

That s 11% higher perf at same power, and 22% lower power at same frequency, the latter amount to 28% higher perf/watt at isofrequency because 1/0.78 = 1.28.

 

[Mahboi]

My 7900XTX runs at 3GHz for compute workloads (pytorch) even when capped at 330W.

I'd love to know how that's possible...someone already mentioned that "RDNA 3 worked well in compute workloads".
But how does the same CU work well in compute but not in render???

 

[Mahboi]

Probably BEOL/FEOL. Less than 2 days for RDNA3.5 to prove that was the case.

BEOL/FEOL?

 

[branch_suggestion]

BEOL/FEOL?

Back end of line/Front end of line.

 

[marees]

I'd love to know how that's possible...someone already mentioned that "RDNA 3 worked well in compute workloads".
But how does the same CU work well in compute but not in render???

My guess is data transfer from mcd to gcd consumes power thus kicking in the power limits

 

[Mahboi]

No, it's not. We've been over this dozens of times.
Navi 33 is a purely monolithic, no MCD/GCD thing, it's the same power problems. It's in the logic.
Same with Phoenix.
Everyone assumes it's the interconnects, it isn't.

 

[Hitman928]

BEOL/FEOL?

@branch_suggestion already provided a visual, but to summarize it, FEOL refers to the substrate, transistors, and contacts to the transistors. BEOL is all of the metal layers and passives (resistors, most capacitors, etc.) in the layers above the FEOL.

 

[Kepler_L2]

I'd love to know how that's possible...someone already mentioned that "RDNA 3 worked well in compute workloads".
But how does the same CU work well in compute but not in render???

Stuff like ROPs/TMUs are not used in Compute.

 

[Mahboi]

Stuff like ROPs/TMUs are not used in Compute.

So the problem would come from the TMUs?
Are the ROPs even that demanding in watts?

 

[Kepler_L2]

So the problem would come from the TMUs?
Are the ROPs even that demanding in watts?

It's not that ROPs/TMUs use a lot of energy, but that they weren't (properly) designed for 3GHz+ clock speeds. So they end up bottlenecking the whole GPU.

 

[Mahboi]

That's so dumb. Just lol.
A whole gen down the drain because nobody remembered to revise the TMUs/ROPs.

 

[marees]

That's so dumb. Just lol.
A whole gen down the drain because nobody remembered to revise the TMUs/ROPs.

The million dollar question is:

Has these been fixed in RDNA 4

& why not release Navi 31 with the 'fixed' design ??

 

[Mahboi]

The million dollar question is:
Has these been fixed in RDNA 4

Yes. No point in releasing it otherwise.

& why not release Navi 31 with the 'fixed' design ??

According to @branch_suggestion, it was IMPOSSIBURI TOO EXUPENSIVU for poor poor AMD. "respins are too expensive".
(I think it's more giving up and penny pinching)

 

[marees]

Yes. No point in releasing it otherwise.

According to @branch_suggestion, it was IMPOSSIBURI TOO EXUPENSIVU for poor poor AMD. "respins are too expensive".
(I think it's more giving up and penny pinching)

Thinking logically, the reason navi 41 & 42 are scrapped is because it is too late to 'fix' it

So the wait continues until RDNA 5 then ????

 

[Mahboi]

I guess that's not totally impossible, but I doubt TMUs or ROPs are the hard part of a GPU anymore, that's more of a 2000s problem. When was the last time anybody cared about the number of ROPs in a GPU?
I think the whole RDNA 4 deboned problem was that it cost them too much. I don't have infos, but I just get a vibe that the product attempted high margins (like always with Absolute Margin Despots) and didn't reach the perf/production costs expectations. So they either fell short in perf, or it was too expensive to make. Most likely the former since the latter should be decently predictable.

They decided to cut all the higher end stuff altogether and ignore GPU while they rake the money with CPUs and focus on CDNA instead. Which yeah, understandable. But boring. Lazy. No-ballz AMD.

 

[SolidQ]

it was IMPOSSIBURI TOO EXUPENSIVU for poor poor AMD. "respins are too expensive".

I remember how they fast fix R520 to R580

 

[Joe NYC]

In the not too distant future they will probably have a much cheaper option from Nvidia with that level of performance, while AMD will have zero until at least 2026.

In other words, you have faith that NVidia will offer better performance / price than AMD?

Good luck with that.

 

[coercitiv]

In other words, you have faith that NVidia will offer better performance / price than AMD?

Good luck with that.

I think his reasoning is Nvidia will offer 4080 Super performance for less than $900. If they keep pumping RT performance, the 7900XTX will look relatively odd even with a hefty discount.

That being said, whatever we get offered in the next 2 years will look like a walk in the park in comparison with the crypto/covid years. Just stock up on keyboards, the ones without a Copilot key will become collector items.

 

[Joe NYC]

I think his reasoning is Nvidia will offer 4080 Super performance for less than $900. If they keep pumping RT performance, the 7900XTX will look relatively odd even with a hefty discount.

4080 Super is competing with 7900 XTX. When 7900 XTX stock runs dry, and there is no competition from AMD, what are the odds that NVidia will lower prices rather than increase prices? We will see.

That being said, whatever we get offered in the next 2 years will look like a walk in the park in comparison with the crypto/covid years. Just stock up on keyboards, the ones without a Copilot key will become collector items.

There is a possibility that the fab capacity tightens in upcoming quarters, and tradeoffs where to channel the capacity will be back, similar to Covid era. It's not a given but a possibility this takes place.

On one side of the tradeoff will be a whiner for whom 7900 XTX was not "good enough", wants more for less from NVidia, on the other hand of the tradeoff, Nvidia can charge $70,000 for datacenter GPU + networking gear at inflated prices.

 

[soresu]

Lazy. No-ballz AMD

It's strategic, and it seems to be paying off as apparently AMD is cutting into nVidia marketshare with US govmt contracts.

If you think it's going to be hard for AMD to take back marketshare with consumer GPUs it's a whole different ballgame with the other pro compute market segments, one which nVidia has been leading very consistently for longer than the consumer side with an ingrained advantage a la CUDA.

So actually it's pretty ballsy indeed, the CDNA roadmap is paying off.

That's not to say that it couldn't shift again if the AI bubble just bursts, but that is talk for another day.

 

[soresu]

I think the whole RDNA 4 deboned problem was that it cost them too much

*Too much for what they can reasonably charge in the consumer market and expect a return.

Even Threadripper is pushing it, and that is largely a replica of the EPYC setup, so it's not costing them a great deal on R&D relative to a super high end consumer GPU.

 

[Mahboi]

*Too much for what they can reasonably charge in the consumer market and expect a return.

Ah great, this is why I write giant articles, so they can cover everything wrong that I hear too often:

AMD is not "lowering costs". They're "maximising margins by lowering costs".

The die is small because Nvidia's ok with outputting a 600mm² giant fat hog of a die. Even if they lose area. Even if it's silly expensive.
AMD's not. So they pour all this effort in magical omega chipletized 144WGP monsters, the monsters all get borked like Frankenstein's monsters and then a manager goes "well, let's just completely give up and go for selling only our tiny monolithic die".
And no, it's not "so utterly borked that it's unsellable". RDNA 3 was sold by AMD, you can trust that RDNA 4 could've been sold above the ridiculous mini die. And you can trust that they cancelled anything above that die because the margins wouldn't have been satisfactory enough for Lisa's lofty penny pinching. It's always don't lose yields, don't take risks, SAVE THE PENNIIIIIIIIIIIIIIIIES...

All the while Nvidia will just make another super expensive 600mm² die.

And do not dare throw a "well, they couldn't", YES THEY CAN.

"Not making huge margins so we won't" is freaking indefensible and I'm not going to hear it. AMD has no more obligations than NVidia to make those margins. They just refuse to make anything less.

Even Threadripper is pushing it, and that is largely a replica of the EPYC setup, so it's not costing them a great deal on R&D relative to a super high end consumer GPU.

Threadripper is a product they should stop making IMO.
I'm convinced there isn't a market for it and that over half the buyers would just take a Ryzen with 40 more PCIe lanes over anything TR.

 

[Hans Gruber]

It's strategic, and it seems to be paying off as apparently AMD is cutting into nVidia marketshare with US govmt contracts.

If you think it's going to be hard for AMD to take back marketshare with consumer GPUs it's a whole different ballgame with the other pro compute market segments, one which nVidia has been leading very consistently for longer than the consumer side with an ingrained advantage a la CUDA.

So actually it's pretty ballsy indeed, the CDNA roadmap is paying off.

That's not to say that it couldn't shift again if the AI bubble just bursts, but that is talk for another day.

If Battlemage is right and the architecture is sound. Intel will be eating AMD's market share big time. I do not understand the logic at AMD. Their drivers are reliable and good. The adrenaline software works well. Nvidia doesn't even have a AI user interface that manages Ocing, fan speeds and provide any kind of useful data.

The company was almost bankrupt before Ryzen. Now all they care about is margins. That is not winning solution long term.

 

[gdansk]

BM needs 2x the performance-per-watt of Alchemist to even be in the same league as the existing 7900 GRE.
And by then N48 will be out. And if it's better what makes you think Intel doesn't care about margins?

 

[Hans Gruber]

BM needs 2x the performance-per-watt of Alchemist to even be in the same league as the existing 7900 GRE.
And by then N48 will be out. And if it's better what makes you think Intel doesn't care about margins?

Because they sold the Arc A750 for $180. Alchemist was based on N6 and supposedly Intel will be using N4P silicon. The good news for consumers. We will have a choice between Intel, AMD and Nvidia all on the 4nm TSMC silicon for the first time ever. That means if a design is efficient, it will not be based on silicon superiority because all three companies will be using the same silicon.

AMD had a huge silicon lead with RDNA1 because it was based on TSMC 7nm which was way ahead of everything else on the market at the time. Nvidia was using Samsung silicon for the 20 and 30 series cards. RDNA1 made people think because the value and performance was there. It was still well behind Nvidia at the time.

Nvidia would have been on 3nm for Blackwell but since N3 sucks other than efficiency. They will wait for the 3nm refresh that will give a performance uplift as well as efficiency gains that was lacking with N3.

If you look at the power consumption of the 4060, it's amazingly efficient. We will find out by early next year if that efficiency was in the design of the GPU or because of the 4nm silicon the 40 series was made on. Blackwell will be based on 4NP which gives much more density for Blackwell than N4 or N4P. I guess Nvidia could not make Blackwell on any other 4nm process. They would have to move to 3nm and TSMC has a capacity issue with 3nm right now. The performance/efficiency should be similar to N4P, but we need TSMC to publish the performance metrics for each respective process.

Both AMD and Intel could be on N4 instead of N4P but then there would be mildly significant performance and efficiency losses over 4NP (Nvidia). For Intel they are going from N6 which was still 7nm while AMD high end GPU's were on N5. So Intel will see bigger gains in efficiency based on silicon alone.