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

 

[SolidQ]

Because there's no market condition where client dgfx ever makes them money back.

they need make top, not everytime cancel etc, or create new uarch

 

[SolidQ]

Real 9070XT score is 26k with release drivers.

We have some info about games, but we don't know, it's real 4080 perf in games or AMD drivers not final yet. Mr.Jebait already leave AMD

 

[adroc_thurston]

they need make top

That's a lotta money.

create new uarch

Well they did. RDNA4 is gfx12. RDNA5 is gfx13.
They gfx IP roadmap is still very aggressive, they're just no longer trying to make discrete GPUs work as a market segment.

 

[SolidQ]

That's a lotta money.

yep, but they did it with Ryzen, why not making with Radeon? They will drop to 5% market share if nothing to do

 

[adroc_thurston]

yep, but they did it with Ryzen

Ryzen offed Intel in DIY DT client with what amounts to server table scraps.

why not making with Radeon?

For graphics, you need to purpose-build everything.

They will drop to 5% market share if nothing to do

Well, yeah. They conceded the market, lmao.

 

[linkgoron]

Of course, RDNA3's issues seem to stem from its architecture rather than its chiplet packaging, but it can be hard to separate the two without direct acknowledgement from AMD on what went wrong.

Even if the chiplet solution wasn't the issue, when you have limited resources (and you always have limited resources) there is an obvious question of how much engineering effort was taken from other parts of the GPU for the chiplet effort. Proponents will say that it wasn't much, or that it was already done with CDNA or whatever - I'd say that RDNA4 chiplet cancellation is proof enough that it was a major issue with why RDNA3 failed.

 

[adroc_thurston]

I'd say that RDNA4 chiplet cancellation is proof enough that it was a major issue with why RDNA3 failed.

RDNA4 and RDNA3 tiling setups are completely unrelated.
just stop.

 

[SolidQ]

For graphics, you need to purpose-build everything.

let's hope maybe UDNA will compete with 6090, or they again concede

 

[adroc_thurston]

let's hope maybe UDNA will compete with 6090

No such thing as UDNA and halo isn't planned.

they again concede

Not again, they more or less soft-exited the market.

 

[SolidQ]

No such thing as UDNA and halo isn't planned.

So another top cancelled?

 

[adroc_thurston]

So another top cancelled?

Eons ago.
Again, they conceded the market.

 

[SolidQ]

Again, they conceded the market.

Then they need new team which can create good product

 

[insertcarehere]

Good news, Meta (the actual real driver of the CUDA roadmap) has probably 10x that, too.

Of which Nvidia actually seems to get that software engineers from Facebook/MS/OpenAI et al are extremely expensive assets whose time is put to significantly better use than debugging BS because of lacking software support from OEMs (no "open source" doesn't do anything if the OEM doesn't step up to the plate). Does not take a lot of extra man hours at $200-300 USD/hour for any upfront system savings to evaporate if adoption comes with more potential issues (and it will, given vast majority of universities with CS grads in related fields train on CUDA - again due to better support).

You can do bigger and better things instead.

Lol because AMD Instinct has been such a great success against Nvidia's products ...as evidenced by the latter's skyrocketing margins, ASPs, Revenues and just about every single financial metric that matters? They won't even have standardized APIs (DirectX) evening the field so all of AMD's shortfalls on software just get put on full display.

 

[adroc_thurston]

Then they need new team which can create good product

They CAN. But they WON'T. Because it's a lot of program cash spent on ??? ROI.
With DC GPGPU it's easy, you sell roadmap as much as you sell the part itself.
But client dgfx is gambling.
Like let's say you've built the chainsaw. You won against NV by like 20% in halo.
Now total of 3.5 cripples bought your part and you basically wasted 4 years to post a major loss.

Of which Nvidia actually seems to get that software engineers from Facebook/MS/OpenAI et al are extremely expensive assets whose time is put to significantly better use than debugging BS because of lacking software support

They still do that lol.
I get it, you only have a cursory understanding of DC things.
Like CUDA s/w stack is good and functional, but that's still GPU software and it breaks. In funny ways too.

Lol because AMD Instinct has been such a great success against Nvidia's products

Yeah? NV is aggressively defending their share and trying to rush GB200 platforms to market.
Don't be naive.

 

[blckgrffn]

Then they need new team which can create good product

More like AI needs to die out (ha) so that AMD and nvidia come back to DGPU as something that could be meaningful. All the best talent @ AMD has to be on server and client/custom client solutions where GPU tech is still needed. That's where the money is.

 

[SolidQ]

But client dgfx is gambling.

They create Ryzen, and now people buying X3D, same can be with Graphic Divison. Only need fix software/media, as we see AMD even Adobe don't want support

 

[adroc_thurston]

They create Ryzen,

server table scraps. Not a purpose-build product.

and now people buying X3D

an even bigger case of server table scraps.

same can be with Graphic Divison

NOPE.
Gotta purpose-build everything.

Only need fix software/media, as we see AMD even Adobe don't want support

irrelevant. things not chainsaw not relevant.

 

[SolidQ]

NOPE.

but anyway jack was saying they will back to halo parts, only question when

 

[adroc_thurston]

but anyway jack was saying they will back to halo parts, only question when

no, they're not going back. Like conceded the market.
He just worded "we surrender" in SVPese.

 

[SolidQ]

no, they're not going back. Like conceded the market.

I'm still think merge ATI and AMD was worst idea. Ati was hell comptetive vs NV

 

[insertcarehere]

They CAN. But they WON'T. Because it's a lot of program cash spent on ??? ROI.
With DC GPGPU it's easy, you sell roadmap as much as you sell the part itself.

DC is so 'easy' which is why AMD has, in nearly its entire existence, been a distant second to Nvidia in anything that is related to professional use?

They still do that lol.
I get it, you only have a cursory understanding of DC things.
Like CUDA s/w stack is good and functional, but that's still GPU software and it breaks. In funny ways too.

'Funny ways' that are easier to predict have better documentation for solutions and are therefore quicker to solve, among other things, because more people have seen similar failure modes which have been better collated (with support), in part that's what building up the substantial software gap between the two companies.

Yeah? NV is aggressively defending their share and trying to rush GB200 platforms to market.
Don't be naive.

The same NV that nearly doubled Revenues with margin expansion on top of that, on a Y/Y basis, with 60% operating margins to AMD's 11%? Where's this 'aggressive defense' cos that sure ain't showing up on a 10K.

 

[adroc_thurston]

DC is so 'easy' which is why AMD has, in nearly its entire existence, been a distant second to Nvidia in anything that is related to professional use

GPGPU wasn't a relevant market until really recently and the moment AMD got some spare change coming GPGPU roadmap came back online.
AMD had like a good 5 year gap between DPFP sticks, after all.

Funny ways' that are easier to predict have better documentation for solutions and are therefore quicker to solve, among other things, because more people have seen similar failure modes which have been better collated (with support), in part that's what building up the substantial software gap between the two companies.

Spoken like someone who never had CUDA setup break in horrendous ways on update.
You're talking in abstract hypotheticals and not anything related to NV devportal experience. Stop.

Where's this 'aggressive defense' cos that sure ain't showing up on a 10K.

Do the channel check and find out.

I'm still think merge ATI and AMD was worst idea. Ati was hell comptetive vs NV

no, ATi was even more doomed.

 

[reaperrr3]

I'm still think merge ATI and AMD was worst idea. Ati was hell comptetive vs NV

R420 vs. NV40 - dud, due to lack of SM3.0 support (NV marketing was already really good at selling mostly irrelevant features)
R520 vs. G70 - dud, too late due to a bug and also not good enough perf/clock when it finally came out
R600 vs. G80 - no need for words

Yes, the ones after were better, but my point is: ATi was hit and miss, not sure they could've survived on their own in the long run as Nvidia already had a clear mindshare advantage at that point.

Also, without GPU revenue and console deals - where AMD was only able to win the contracts due to having both decent small x86 cores and decent GPU tech - AMD as CPU-only company would've probably died during the Bulldozer era, and there would've been no Ryzen.

Basically, we'd probably live in an Intel x86 monopoly + NV GPU monopoly world since years ago if that merger hadn't happened.

 

[SolidQ]

R420 vs. NV40 - dud, due to lack of SM3.0 support

SM 3.0 was in like 2-3 games, Far Cry HDR and splinter cell and maybe yet one game was

R520 vs. G70

R520 yes was bad, but R580 easily beat any G70

R600 vs. G80 - no need for words

Here nothing to say, but HD 5/6/7 was very good

 

[inquiss]

Agree, the issue is the rumored $599 price tag which does barley nothing once again for generational performance/dollar uplift. you could get a GRE for that price some weeks ago. (actually still now available for that price in my local market and currency adjusted and it is actually then $599 incl. all taxes, while with the $599 MSRP + taxes plus "EU tax" it will end up with worse performance/$ than a 7900 GRE)

I find this an odd argument. A new product is about to come out, so the old product with similar performance gets discounted to what the new product will offer, and when it releases people say it doesn't increase value.

I'm not trying to defend or really comment on the suggested prices for these new cards, more a comment on the thinking that heavily discounted previous gen is the best comparison.