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

 

[moinmoin]

TruForm was just an experiment.

Everything is an experiment until it isn't anymore.

 

[GodisanAtheist]

It's shocking to me that we're not getting two gens of RDNA3 style chiplets.

It's a good (great) idea. It works. Now AMD just has to go balls out with it. Give us that big GCD.

RDNA3, with all of its flaws, would be plenty competitive with AD102 if they just scaled up to 400mm2+

Plenty of stopgap there for them to figure out

 

[adroc_thurston]

Or will they go back to something closer to RDNA3.

no.

RDNA3, with all of its flaws, would be plenty competitive with AD102 if they just scaled up to 400mm2+

That wasn't the idea.

 

[jpiniero]

no.

So what are they going to do if they can't get it to work? Cancel that too?

 

[adroc_thurston]

Cancel that too?

Yea.

 

[Ajay]

Yea.

One can only hope that RTG learned enough from the mutli-chiplet RDNA4 to hit their targets on RDNA5. Halo GPUs help sell downstream products.

 

[adroc_thurston]

RTG learned enough from the mutli-chiplet RDNA4 to hit their targets on RDNA5

This isn't anout targets.
Making the device itself work correctly is a pickle.

 

[Kepler_L2]

One can only hope that RTG learned enough from the mutli-chiplet RDNA4 to hit their targets on RDNA5. Halo GPUs help sell downstream products.

The biggest reason for making a Halo GPU atm is to fill in the gap below the true AI/HPC behemoths like MI300/400 and GB/GR100.

 

[Ajay]

This isn't anout targets.
Making the device itself work correctly is a pickle.

TTM targets, so they don’t miss the release window again.

 

[adroc_thurston]

TTM targets, so they don’t miss the release window again.

Well yea but that applies to every single chip ever made.

 

[Ajay]

The biggest reason for making a Halo GPU atm is to fill in the gap below the true AI/HPC behemoths like MI300/400 and GB/GR100.

Yes, that aspect has changed over the past decade.

 

[Ajay]

Well yea but that applies to every single chip ever made.

True. Also true for all complex high tech products.

Kinda of OT, but I’ve been there when projects start getting cancelled. People start jumping ship - I jumped to another engineering group in a different city. Aside from recessions, there's always a talent shortage. Maybe AMD RTG has more loyal engineers.

 

[igor_kavinski]

Maybe AMD RTG has more loyal engineers.

Where else they gonna go? If they were any better, they would already be working at Apple or nGreedia. Intel won't hire them unless they come up with a guaranteed RTX beating design.

 

[Aapje]

With the long lead times from initiating, designing and releasing new chips the changes in graphics APIs and their adoption in games while slow actually can happen faster than actually new chips getting launched. So it's not unfathomable that AMD originally bet on graphics APIs having progressed significantly more by now, and that having turned out to be a wrong prediction, N41 and N42 thus having been designed for a reality that never came to be.

We are overdue for a DirectX 13 release, which is kinda crazy, given the extent to which DX12 delivered on its promises. Normally you'd expect a make-up release relatively quickly.

 

[igor_kavinski]

We are overdue for a DirectX 13 release, which is kinda crazy, given the extent to which DX12 delivered on its promises. Normally you'd expect a make-up release relatively quickly.

Gonna have to be a Win12 feature, using AI acceleration.

 

[eek2121]

That would essentially kill N32 cards, do you believe AMD would do such a bold move?

As they killed N22, etc.

It would be right in line with AMD's history of self destructive GPU business strategies. G80 wasn't the beginning of the end for AMD's aspirations of graphics leadership, RV770 was. AMD needed a R600/G80 moment, and could have had it just one generation after G80, but despite a far superior architecture decided to only take a midrange die to market. AMD's GPU side has always had issues delivering the right product at the right time. You can't really predict performance based on what segments they are releasing to until AMD's history of incompetent GPU business decisions is sufficiently in the rear view mirror.

That said, the narrative going around is that high end RDNA4 was shelved because AMD bit off more than they could chew with advanced packaging, and to the extent that is true the decision to only release low end parts may have nothing to do at all with performance projections, and everything to do with not having monolithic backup plans in place for those segments.

A lot of folks are confusing ultra premium titan killer wannabe with mid range . AMD will have multiple competitive offerings next gen.

N21 worked as intended, N32, like the rest of RDNA3 has not.

AMD is launching a pro variant of cards that uses N32. N32 is doing just fine.

IMO, they can't release a GPU that doesn't work with DX11 or OpenGL or whatever. It has to work like a monolithic GPU.

There are solutions to this. Small monolithic backwards compatible chip for DX 9-12, new chip for new stuff. Of course, nobody will do this because 💰💰💰

Be kind of hard to sell a gaming GPU without graphics.

Ehhh…3DFX? 🤣

We are overdue for a DirectX 13 release, which is kinda crazy, given the extent to which DX12 delivered on its promises. Normally you'd expect a make-up release relatively quickly.

It is kinda in the works.

 

[moinmoin]

We are overdue for a DirectX 13 release, which is kinda crazy, given the extent to which DX12 delivered on its promises. Normally you'd expect a make-up release relatively quickly.

Indeed. One aspect that's really overdue is a more flexible actually scalable RT API that's usable on weaker hardware as well. The issue appears to be nobody is interested in pushing such right now. Nvidia is obviously fine with keeping status quo pushing its huge highest end monster cards. AMD is in no market position to push for a sustainable change in RT APIs. Microsoft likely waits for the next console gen before introducing any change to Windows. Other players are mainly still catching up to the status quo.

 

[adroc_thurston]

We are overdue for a DirectX 13 release

DX12u was effectively that.

A lot of folks are confusing ultra premium titan killer wannabe

That's the only way to sell any GPUs yes.

 

[branch_suggestion]

AMD is launching a pro variant of cards that uses N32. N32 is doing just fine.

Well yeah, it functions fine and is still an improvement over RDNA2, but perf/power targets were missed badly.
RDNA3 is still cost competitive, but AMD had to take a chunk off the margins they actually wanted this gen.

 

[Strangerer]

https://www.amd.com/en/corporate/events/advancing-ai.html

 

[Joe NYC]

Cheese tweeted a picture of Gigabyte rack unit with 4x Mi300a (CDNA3)

https://twitter.com/x/status/1724491521460785162

 

[Mopetar]

Actually quite interesting. Any idea if AMD starts planning to offer 4P solutions for Epyc as well, or is this just for these parts specifically?

 

[Joe NYC]

Actually quite interesting. Any idea if AMD starts planning to offer 4P solutions for Epyc as well, or is this just for these parts specifically?

AMD is leaning in favor of 1P for Epyc, and adding extra core to that single processor.

Also, AMD was so concentrated on getting the El Capitan finished on time and Mi300x to market that the CPU solution based on Mi300 were likely on a side track. But we will see what 2024 brings...

 

[igor_kavinski]

An EPYC Miss? Microsoft Azure Instances Pair AMD's MI300X With Intel's Sapphire Rapids

Genoa sidelined by Sapphire Rapids.

www.tomshardware.com

Possibly volume or pricing issue that prevented Microsoft from going Epyc.

 

[Joe NYC]

An EPYC Miss? Microsoft Azure Instances Pair AMD's MI300X With Intel's Sapphire Rapids

Genoa sidelined by Sapphire Rapids.

www.tomshardware.com

Possibly volume or pricing issue that prevented Microsoft from going Epyc.

Could also be a timing issue.

NVidia probably offers a reference design for the 8x H100 that is using Sapphire Rapids. Maybe Microsoft found out that they can just swap the cards with 8xMi300x and everything just worked. So no reason to start validating a new platform, since everyone is racing to deployment.