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

 

[Mopetar]

Yeah, even if it were a bubble, they'll take the profit and use it to R&D in preparation for the next hype wave, be it self driving cars or quantum computing. Rinse and repeat.

It hardly matters to NVidia whether it is or not. They're just going to keep selling as much as they can and the only way they get burned is with excess inventory. However, the margins are so good that once they see the end coming they can start cutting prices to sell through whatever they have while still making a lot of profit.

There may be some investors who get burned if the stock price goes further through the roof. It might already be there for all I know. But that's not going to hurt NVidia at all. Even if the AI market does collapse at some point, it's still possible that they've got products to catch the next wave. JHH certainly has a good record for knowing where the market is headed and delivering on that. Even if he gets it wrong he can still try to create the perception of having the next big thing as NVidia usually gets to where it wants to be ahead of their competitors.

 

[adroc_thurston]

Around end of 2024 to be the top in revenue.

Unlikely, current GPGPU spending bump is mostly due to a massive pull-in by the likes of China in particular.

 

[Joe NYC]

Unlikely, current GPGPU spending bump is mostly due to a massive pull-in by the likes of China in particular.

Interesting point. So if there is a pull in, of demand, to beat the sanctions, then, there is going to be a reduction from China, once all the stupid sanctions kick in.

In the meantime, Huawei is claiming to have an A100 level home grown card.

 

[adroc_thurston]

once all the stupid sanctions kick in.

They're not stupid.
Very reasonable.

Huawei is claiming to have an A100 level home grown card.

Chinesium vendors make monthly claims but they're never ever real.
HiSilicon track record with anything resembling homegrown IP is disasterous.

 

[Ajay]

Oh it is.
I'll give it 4-6Q before startups go mass extinction/exit scam.

Startups - yes. NV, don't think so. We'll see if any of the well funded companies like Alphabet, or Intel can do anything - probably not.
It's not a scam, just overblown. ML has a bright future, AI is the hype.

 

[adroc_thurston]

NV, don't think so.

NV directly relies on startups and all other flavours of solutions searching of a problem stimulating H100 demand.
When all those solutions find no problem to solve, poof, demand's gone.

 

[Ajay]

NV directly relies on startups and all other flavours of solutions searching of a problem stimulating H100 demand.
When all those solutions find no problem to solve, poof, demand's gone.

Well, I'd have to see the numbers. There are large and medium size companies building the infrastructure for 'AI'. What they'll really be doing is building large ML systems, which will pay off. AI is to hype up consumers and become a buzzword that companies are racing to add to their names (just saw google proffering itself as an 'AI company" - lol). The application space for ML via DLN (DNN) is growing and so will the TAM. Inferencing hardware is already in many consumer devices and will be ready to take advantage of cloud based system capability. Advances in hardware and software, particularly by NV are enabling these related nascent technologies to finally hit stride.

That said, if the numbers prove it's mainly hype - I'll be okay with that, I don't have any skin in the game.

 

[adroc_thurston]

There are large and medium size companies building the infrastructure for 'AI'.

It's just C-listers trying to ride the hype wave.

The application space for ML via DLN (DNN) is growing and so will the TAM

Solutions, yet they solve no problem.
None of it has a business model.

Advances in hardware and software, particularly by NV are enabling these related nascent technologies to finally hit stride.

Solutions without problems or a business model.

 

[Joe NYC]

It's just C-listers trying to ride the hype wave.

Solutions, yet they solve no problem.
None of it has a business model.

Solutions without problems or a business model.

Precisely.

If the revenue is from the datacenter graphics cards is going to be 100 billion, there has to be 10x that, or 1 trillion revenue generated by the buyers of these cards, and of that, 100 billion profit to be able to afford next years' worth of cards.

I don't see a trillion in new revenue being generated...

 

[Ajay]

It's just C-listers trying to ride the hype wave.

Solutions, yet they solve no problem.
None of it has a business model.

Solutions without problems or a business model.

Well, I can't (quickly) find any useful customer list - aside from a few NV cherry picked on their own website. So no data to support any claims.

I understand there needs to be a business model. I think that there are many businesses that rely on software already, where the business case will be improved performance, customer responsiveness, new features, etc. So it will be more of a dovetail effect. New 'AI' businesses, yeah, there's a bunch of startups making the rounds in business talk shows like on CNBC touting the promise of their wares. Maybe they'll all fold, maybe a couple will actually create a new market (too soon to know). I think you are overly confident that all things 'AI' will fail; that it's all a scam - I disagree. I think there will be additive effects in existing companies that depend on analytics involving very large datasets and/or very complex tasks (like making multi-billion xtor semiconductor chips). Oil & Gas, pharmaceuticals are all companies already using super computers - creating training models that more quickly discard outliers or non-productive paths/fields would be worth huge $$$s.

Anyway, I guess we will see within the next two years. What has the stock market priced NV at - 2-3x forward looking profits; not 10-20x. So money managers probably aren't going for all the hype, though I wonder if small investors are.

 

[Frenetic Pony]

Anyway, I guess we will see within the next two years. What has the stock market priced NV at - 2-3x forward looking profits; not 10-20x. So money managers probably aren't going for all the hype, though I wonder if small investors are.

Nvidia is still the most overpriced stock I've seen in a long while. Just another frontman for ridiculous hypeway, same as any other. The idea that they somehow have a sustained advantage in what is ultimately a lesser HW spend, AI training vs AI deployment (deployment is where far more volume could be) is already ridiculous. With diverging AI paths there's likely to be diverging "winners" for training to begin with, not just one, and certainly not 1 that can sustain the kind of margins Nvidia is currently seeing.

The idea of GPUs being used for inference is interesting though, I mean the math is similar and if the HW is already there, then... sure maybe? Especially if it's a more "workstation" type thing and not a small laptop/phone SOC. I could see GFX5 supporting full utilization float8, the new inference hotness being pushed. I already use AI for my photography hobby, both DXO Raw denoise and Photoshop upscaling, it's fine at the current speed but I can only imagine how slow it might be for a similar quality doing video. Same goes for say, making rendering assets using generative AI, already a very real though very early market.

I.E. Generative AI is, and already does, have a very real market. Helping train low level types in tech support or automating basic dumb answering services or even generating graphics for business presentations/internet articles, or etc. None of this is nearly as profitable or exciting as dumb tech bros make it out to be, but then that's how the hype wave usually goes.

 

[Joe NYC]

Well, I can't (quickly) find any useful customer list - aside from a few NV cherry picked on their own website. So no data to support any claims.

I understand there needs to be a business model. I think that there are many businesses that rely on software already, where the business case will be improved performance, customer responsiveness, new features, etc. So it will be more of a dovetail effect. New 'AI' businesses, yeah, there's a bunch of startups making the rounds in business talk shows like on CNBC touting the promise of their wares. Maybe they'll all fold, maybe a couple will actually create a new market (too soon to know). I think you are overly confident that all things 'AI' will fail; that it's all a scam - I disagree. I think there will be additive effects in existing companies that depend on analytics involving very large datasets and/or very complex tasks (like making multi-billion xtor semiconductor chips). Oil & Gas, pharmaceuticals are all companies already using super computers - creating training models that more quickly discard outliers or non-productive paths/fields would be worth huge $$$s.

Anyway, I guess we will see within the next two years. What has the stock market priced NV at - 2-3x forward looking profits; not 10-20x. So money managers probably aren't going for all the hype, though I wonder if small investors are.

A good indicator is how many 100s of billion dollars have already been invested to put drivers (taxi drivers, truck drivers) out of work. And their pesky children who want to eat, want shoes, clothes, maybe computer to play some games.

After all this has been invested to destroy the truck driver, the truck driver is still standing while still more, additional 100s of billions are pouring in to ruin his life.

 

[Tigerick]

RGT mentioned that RDNA4 GPU could launch first half of next year; that is pretty aggressive timings. Is it possible AMD could launch mobile version first to replace current version N32/N33 especially mobile N32 is MIA??

 

[coercitiv]

RGT mentioned that RDNA4 GPU could launch first half of next year; that is pretty aggressive timings.

The same RGT who said N33 is "very very power efficient" and N31 is over 2X performance of N21? When will people stop lending their ear to these channels?

 

[SteinFG]

RGT mentioned that RDNA4 GPU could launch first half of next year; that is pretty aggressive timings. Is it possible AMD could launch mobile version first to replace current version N32/N33 especially mobile N32 is MIA??

Considering that Navi 4X, or whatever its codename, is not expected to beat Navi 31, there is a small chance for what you are saying.

Unrelated guess: AMD likes simple scalable designs (N10 is N14x2, N21 is N22x2, N23 is N24x2, N31 is almost N32x2), I think that Navi 4X is Navi4M doubled, maybe they are 60 and 30 cu, or 60 and 36, or 72 and 36. That's assuming 3 SA per SE.

 

[PJVol]

The same RGT who said N33 is "very very power efficient" and N31 is over 2X performance of N21? When will people stop lending their ear to these channels?

There were those who claimed otherwise back then? Remind me please, just to occasionally lend an ear.

 

[coercitiv]

There were those who claimed otherwise back then? Remind me please, just to occasionally lend an ear.

So just because everyone was regurgitating the same fake leaks makes RGT more credible? Maybe you should lend your ear to those who chose to stay silent in the absence of verifiable information, it's still not great practice but an improvement nonetheless.

 

[SolidQ]

There were those who claimed otherwise back then?

sometimes they saying it's TF perfomance, not gaming

 

[PJVol]

So just because everyone was regurgitating the same fake leaks makes RGT more credible?

That doesn't make him any less, does it?

Maybe you should lend your ear to those who chose to stay silent in the absence of verifiable information

Come on .... 🤣

 

[PJVol]

sometimes they saying it's TF perfomance, not gaming

If you asked me, I'd confirm that too, TF is not FPS.
Just remind me did "they" mention dual issue SIMDs?

 

[Ajay]

If you asked me, I'd confirm that too, TF is not FPS.
Just remind me did "they" mention dual issue SIMDs?

The dual issue SIMDs aren't really straight up dual issue. The ability to dual issue is limited to certain conditions. I hope RDNA5 improves on this. NV has had a well functioning dual issue ALU unit for a while.

~70% of game code if FP and ~30% is INT - NV uses one FP plus a mixed FP/INT unit (IIRC) to cash in on this fact.

 

[PJVol]

The dual issue SIMDs aren't really straight up dual issue. The ability to dual issue is limited to certain conditions. I hope RDNA5 improves on this. NV has had a well functioning dual issue ALU unit for a while.

~70% of game code if FP and ~30% is INT - NV uses one FP plus a mixed FP/INT unit (IIRC) to cash in on this fact.

Thanks for the info, but this wasn't the point of my question.

 

[CakeMonster]

There's no way AI and ML is a bubble, those have several use cases already and I'm sure more will be invented, utilizing current hardware. But the startups with bullsh*t artists who obviously are just using existing (or merged) LLM's or Stable Diffusion are a sure sign of hype. Everytime hype riding CEO's or WSB or regular joes let themselves be fooled into investing because they haven't bothered understanding the current state of a market or technology, there is an eventual reckoning. It might not bite NVidia stock hard, but it will sure have a market impact.

 

[RTX2080]

An unknown amd (discrete) GPU on N4 had been sent to factory for early testing few days ago. Not production purpose.
This post is just for a record and might be testified months later, or might be not.

Now confirmed it is RDNA4. Possible release timing window could between Q2-Q3 2024.

 

[Aapje]

With pretty much any new technology there will be failed attempts to monetize it, which usually will keep getting financed by investors who buy into the hype and don't apply critical thinking. Then something will damage that trust and suddenly investors will panic and start to demand that these companies demonstrate that they can become profitable. Then the good survive and the bad get culled. It's pretty much a given that this will happen to AI too, because investors will keep making the same mistakes.

Of course, when that happens, the doom sayers here and elsewhere will cry victory and will write off the technology, making the exact opposite mistake of the investors, ignoring all the good, just like the investors at first ignore all the bad.

For example, we had the exact same thing with the dotcom-hype/bubble, which resulted in a spectacular crisis, but the (commercial) internet was not at all a failure. It replaced many things that used to be offline, or became complementary to offline things. For example, Walmart wouldn't dream of closing their website.

As for the actual use cases, many seem rather obvious. Programmers are already using code generators, people let it write articles and papers (including at my job), newspapers are experimenting with letting the AI make summaries (I am quite confident that we will get Twitter-bots that newspapers will use to generate tweets for articles, which will be trained to produce the right amount of clickbait), people are already winning art competitions with AI (and surely we'll be generating art for presentations soon). Search engines, which somehow have been getting worse and worse over time, will almost certainly get replaced by AI. Personal assistents are likely. Lonely people (at first the elderly) will get AI friends (I know a company who has a rudimentary form of this), etc.

But ultimately we almost always see innovation beyond what we can predict, and it seems rather predictable that since AI has so many possible ways to use it, that we'll see things be a success that most of us now can't even fathom.