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

 

[linkgoron]

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

Good luck with that.

They don't need to offer better price/performance than AMD. They will offer a performance level that AMD does not provide. Higher end cards always provided worse price/perf than lower-tier cards - and at the high end It's either Nvidia or nothing. 5090 will be WAY out of AMD's reach, and the 5080 will probably be at least at 4090 level (note that the 4080 is actually ~12% faster than the 3090 TI and ~30% faster than the 3090). 4080 Super was released relatively recently and is $1000 MSRP, and while it's slightly slower in raster than the 7900xtx it's got better efficiency and much better RTX. $1000 vs $900 isn't that big of a difference.

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.

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.

4080 was $1200 and 4080 Super has a $1000 MSRP. If the 5xxx is a relatively strong series, Navi 48 might be competing with 5060TI or at best with the 5070. Nvidia could release a 5080 that'll be a bit better (say, 10%) than the 4090 at $1200 or $1300 and people will buy it.

 

[linkgoron]

Gamers still pumping hopium for someone making their green slop cheaper.

People keep saying this, but Nvidia's execution has been terrific (except for maybe the 30 series, which was IMO mediocre) and AMD has been mediocre to bad. Inconsistent naming, inconsistent releases etc. I'm pretty sure that people said that people wouldn't buy AMD vs Intel as well, and that people want AMD for cheap Intel. Once the performance was there people bought AMD. Yes, Intel are not Nvidia and they destroyed themselves and were there for the taking, but it's not like people want better AMD stuff to get cheaper Intel.

AMD haven't taken the performance crown in more than 10 years, basically Nvidia has had better performance since Maxwell. Sometimes significantly. It's not like AMD have had these great performing cards generation after generation and people are just buying Nvidia anyway. AMD basically had Polaris and somewhat Vega for years.

Yes, stuff has turned around since RDNA (I still have a 5700xt myself), and AMD usually provides somewhat better price/perf (although not always significantly better), but they still have really weird decisions like last-minute price cuts and just weird nonsense like the 7600 and bad pricing like the 7900XT. Nvidia provides a lot of things to the "casual" user. Yes, they've got a strong branding but they've also got relatively consistent branding and naming (except for that weird 4080 12GB card that they unlaunched), and timely top-to-bottom releases and refreshes. AMD hasn't had any of those in years those except for RDNA2, which was the only really competitive generation since forever. RDNA3 was a mess, and the high-end of RDNA4 has been completely cut.

 

[blackangus]

There is that old saying. We live in interesting times. My answer, no not really.

This is not twitter. You are not limited in characters and you are allowed to explain your point of view in depth rather than staying with throwaway one-liners.

Well to Hans point, its not a simple as either of the sides he or I was taking. Both were lopsided statements and like everything the truth is in the middle with lots of complexity.
If you take a snapshot and only talk about NOW, then AMD needs care alot about margins, because they do need to keep their RnD going strong.
As market conditions change it is possible there would be a place where AMD could accept lesser margins w/o hurting their RnD in a way that would be very detrimental.
But that is the future, and way to foggy for me to call!

 

[SolidQ]

Interesting, will RDNA4 use with VCN5 Xilinx AV1, to beat NV encoder?

 

[soresu]

Interesting, will RDNA4 use with VCN5 Xilinx AV1, to beat NV encoder?

Whatever it is I hope it's better than what was in RDNA3.

I frequent an AV1 encoding discord and the impression of encoding quality from that HW ASIC is very low.

 

[adroc_thurston]

I frequent an AV1 encoding discord and the impression of encoding quality from that HW ASIC is very low.

Q is fine, but it has hilarious errata.

 

[Mahboi]

Whatever it is I hope it's better than what was in RDNA3.

I frequent an AV1 encoding discord and the impression of encoding quality from that HW ASIC is very low.

frankly these types can be very anal about things.
I've recorded AV1 with RDNA 3, it's very solid. The complaints I heard was from people who essentially expected lossless, which is insane to expect out of HW.
The only real BIG problem is that recording/streaming 4K with RDNA 3 was somewhat beyond what it could do. Like the media engine would get juuuuuust overloaded, but if you lowered to something like 3400x2000 from 3840x2160, it was fine.

 

[soresu]

Like the media engine would get juuuuuust overloaded, but if you lowered to something like 3400x2000 from 3840x2160, it was fine

At what FPS?

 

[Mahboi]

At what FPS?

Monitor 144, recording 60.

 

[soresu]

Monitor 144, recording 60.

From what I heard the HW decode for AV1 also suffered frame drops at 8K, tho I can't remember the FPS in question.

 

[Joe NYC]

They don't need to offer better price/performance than AMD. They will offer a performance level that AMD does not provide.

$1,800 for 4090 is a good preview of what happens in the situation you describe.

 

[Joe NYC]

So, looks like nothing related to GPU during the keynote.

Either that or Lisa was just too embarrassed to link @AMDGaming Twitter account.
(and that account it is an embarrassment)

But @AMDInstinct included, so maybe some CDNA4 info

https://twitter.com/x/status/1797276920331256224

 

[inquiss]

So, looks like nothing related to GPU during the keynote.

Either that or Lisa was just too embarrassed to link @AMDGaming Twitter account.
(and that account it is an embarrassment)

But @AMDInstinct included, so maybe some CDNA4 info

https://twitter.com/x/status/1797276920331256224

Think it's more likely the revised version of MI350X but hope it is something more substantial

 

[tsamolotoff]

similar to R300 + release earlier = more market share.

N2x was of the same sort of PPW and PPD win as, lets say, Cypress and RV870 (well, hawaii too), yet it was still outsold like 10 to 1 (especially low end models which were both faster and cheaper than competitor options)

 

[Joe NYC]

Think it's more likely the revised version of MI350X but hope it is something more substantial

I think they will have a roadmap with Mi400 on it, even if there are no details.

 

[Abwx]

I think they will have a roadmap with Mi400 on it, even if there are no details.

AMD just confirmed these days 4bn $ shipment of MI300 for 2024, given the current AI frenzy they can easily double this number in 2025 if they have some technical advantage, that would surpass their servers or CPU client sales within 2 years.

Guess that s where they ll put their priorities, gaming GPUs will get the minimalist treatment with mid range products that cover 80% of the market, not a bad strategy after all.

 

[adroc_thurston]

Guess that s where they ll put their priorities, gaming GPUs will get the minimalist treatment with mid range products that cover 80% of the matket, not a bad strategy after all.

wrong, Navi50 is still coming.

 

[blckgrffn]

AMD just confirmed these days 4bn $ shipment of MI300 for 2024, given the current AI frenzy they can easily double this number in 2025 if they have some technical advantage, that would surpass their servers or CPU client sales within 2 years.

Guess that s where they ll put their priorities, gaming GPUs will get the minimalist treatment with mid range products that cover 80% of the market, not a bad strategy after all.

Yeah, I think this is just reality of the times.

As long as we get something good to tide to push the competitive envelope in terms of pricing and all that, I guess that's all the win we are getting.

 

[blackangus]

wrong, Navi50 is still coming.

From your lips to gods ears.

 

[Abwx]

wrong, Navi50 is still coming.

Possible, but that was surely on the drawing boards well before the AI market surged exponentialy, so that cost not much to release what was already well advanced last year.

Yeah, I think this is just reality of the times.

As long as we get something good to tide to push the competitive envelope in terms of pricing and all that, I guess that's all the win we are getting.

The success of the 7800XT is a testimony that there s more sales around 500-600$, not that something that is faster is not necessary but above 1000$ there s not much people left, most are die hard Nvidia buyers that are willing to eat potatoes all the year to buy their brand s fastest card, not so much with the AMD buyers who are more price conscious.

 

[adroc_thurston]

From your lips to gods ears.

so true.

Possible, but that was surely on the drawing boards well before the AI market surged exponentialy, so that cost not much to release what was already well advanced last year.

GPGPU stuff does not really directly affect client gfx roadmap.

 

[Abwx]

GPGPU stuff does not really directly affect client gfx roadmap.

They share the underlying designs, they use the same ALUs FI, it s what is around that is adapted to the different tasks, FI rendering back end/GDDR controler for client and HBM controler for GPGPU that do not need the rendering part.

 

[adroc_thurston]

They share the underlying designs

no?
Very different CUs now.

FI rendering back end/GDDR controler for client and HBM controler for GPGPU that do not need the rendering part.

No, they're distinct microarchitectures.

 

[Abwx]

no?
Very different CUs now.

No, they're distinct microarchitectures.

That s still 2 ops/cycle/ALU since ages, at least for 32b and 64b FP, granted that the format of the operands have been extended from 32b/64b to 16b, or even INT8, and the exe units surely splitted to double the throughput when the required precision is halved.

 

[adroc_thurston]

That s still 2 ops/cycle/ALU since ages

Well yeah that's what fused multiply-add is.
FLOPS are only part of what makes a GPU. see Intel gen12 parts for more. or Vega.