Discussion RDNA4 + CDNA3 Architectures Thread

DisEnchantment · Mar 23, 2022

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

Tuna-Fish · Nov 16, 2024

Even if they match it on compute, they need to perform miracles to get the performance they need out of their narrow and slow memory interface.

gdansk · Nov 16, 2024

Tuna-Fish said:
Even if they match it on compute, they need to perform miracles to get the performance they need out of their narrow and slow memory interface.

Or use faster memory than rumoured
It wouldn't be the first time rumors are wrong (although rumors about AMD cheaping out on memory interface have an almost 100% track record...)

poke01 · Nov 16, 2024

I see only Nvidia taking GDDR7 shipments. who knows AMD did as well and we missed it.

we will see how GDDR7 impacts real world apps and games. But the bandwidth increase is really high.

marees · Nov 17, 2024

Tuna-Fish said:
Even if they match it on compute, they need to perform miracles to get the performance they need out of their narrow and slow memory interface.

So what is the limit to n48xtx performance given the memory bandwidth constraints?

SolidQ · Nov 17, 2024

marees said:
So what is the limit to n48xtx performance given the memory bandwidth constraints?

i remember some post 5 years ago about RX 5700XT, where was not very bandwidth dependence. Would be interesting tests with RDNA2/3 dependence

ToTTenTranz · Nov 18, 2024

Tuna-Fish said:
Even if they match it on compute, they need to perform miracles to get the performance they need out of their narrow and slow memory interface.

Memory bandwidth requirements go down if the GPU relies on aggressive upscaling technologies (meaning lower base resolution). One such example is the PS5 Pro PSSR patches in games using lower base resolution than the original PS5 versions.

If FSR4 works as well as PSSR (and there's a good chance one is just a fork of the other), then we should get decent 4K output out of 1080p base resolution.

igor_kavinski · Nov 19, 2024

TensorWave plans world’s largest AMD GPU cluster with gigawatt capacity by 2025 | igor´sLAB

TensorWave, a cloud service provider focused on AMD’s AI technology, has announced that it is developing GPU clusters based on AMD’s Instinct MI300X, MI325X and MI350X accelerators. The aim is to…

www.igorslab.de

Jensen may need to sit down before reading that.

MrTeal · Nov 19, 2024

I'm sure once he gets done swimming laps in his gold vault Scrooge McDuck-style he'll take a break to wipe away the tears.

SolidQ · Nov 19, 2024

2026? isn't that was rumor RTX 6xxx also 2026?

AMD UDNA "Radeon" Gaming GPUs Rumored To Enter Mass Production In Q2 2026, Sony PS6 Also Expected To Utilize Next-Gen Architecture

AMD's UDNA architecture-based Radeon gaming GPUs will allegedly enter mass production in Q2 2026 & also be integrated into Sony's PS6.

wccftech.com

There is no RDNA5 code name. After RDNA4, it will be UDNA.
MI400 and RX9000 use the same UDNA, and the architecture uses an ALU design similar to GCN.
UDNA Gaming GPU is tentatively scheduled for mass production in Q2 26.
Sony's PS6 will use UDNA, and the CPU has not yet been determined to be ZEN4 or ZEN5. Sony's handheld will also use AMD hardware.
I heard that Microsoft's handheld will choose between Qualcomm and AMD? I don't know about this
The above information comes from the supply chain, and I don't know the specifications and performance.

ToTTenTranz · Nov 19, 2024

SolidQ said:
MI400 and RX9000 use the same UDNA, and the architecture uses an ALU design similar to GCN.

I find this a bit hard to believe, considering how much of a jump in performance-per-area RDNA brought in comparison to GCN.

SolidQ · Nov 19, 2024

ToTTenTranz said:
I find this a bit hard to believe, considering how much of a jump in performance-per-area RDNA brought in comparison to GCN.

diffrent times, different teams, diffrent money

soresu · Nov 19, 2024

SolidQ said:
MI400 and RX9000 use the same UDNA, and the architecture uses an ALU design similar to GCN.

Seems unlikely.

CDNA concentrated almost exclusively on improving specific aspects of the architecture around the new matrix units and AI/ML compute.

Doing this allowed them to keep CDNA1-4 more or less consistent in programming model

RDNA meanwhile has not simply added RT and changed to wave32.

It has evolved significantly at the base level of the µArch, and many of the changes in RDNA3 do not seem to favor gaming compute much at all.

In fact I'm inclined to think that RDNA3 was in fact setting the stage for bringing professional compute and gaming back together.

It would surprise me if GFX13 or whatever comes next is basically just GFX9 on steroids.

del42sa · Nov 20, 2024

soresu said:
It would surprise me if GFX13 or whatever comes next is basically just GFX9 on steroids.

if that´s the case, then VEGA - like gaming GPU is back and this means no moving forward, it´s actually a step back ...

adroc_thurston · Nov 20, 2024

soresu said:
It would surprise me if GFX13 or whatever comes next is basically just GFX9 on steroids.

Lmao no, client and DC will forever be almost unrelated.

SolidQ · Nov 20, 2024

AI AI AI AI

Gideon · Friday at 3:31 AM

soresu said:
It has evolved significantly at the base level of the µArch, and many of the changes in RDNA3 do not seem to favor gaming compute much at all.

This!

I find the dilemma whether it uses Zen 4 or Zen 5 equally odd. Why would it use Zen 4? By that point Zen 5 has been out for 2 years. If the want to skimp on area they can just use Zen 5c and cut it down further if needed, as they did with zen 2 in PS5 (128-bit FP unit only)

Gideon · Friday at 5:49 AM

SolidQ said:
MI400 and RX9000 use the same UDNA, and the architecture uses an ALU design similar to GCN.

A respected graphics engine developer also calls BS on this one.

https://twitter.com/x/status/1859879410196955434

I'll quote the whole thread (highlights mine):

Sebastian Aaltonen said:
AMD UDNA will be interesting.

CDNA3 architecture is still based on GCN 4 cycle wave64 scheduling. RDNA schedules every cycle and exposes instruction latency. Scheduler runs/blocks instructions concurrently/dynamically. RDNA is much closer to Nvidia GPUs.

Thread...

CDNA has wide matrix cores and other wide compute workload improvements, which AMD wants to bring to UDNA. It also has multi-chip scaling.

Rumors tell that RDNA4 will finally have matrix cores in consumer space. Seems that AMD is integrating matrix cores early to RDNA lineup.

My expectation is that UDNA compute unit will be RDNA4 descendant instead of CDNA3 descendant. They definitely need 1 cycle low latency scheduling in consumer space, and Nvidia does well with it in AI space too. I don't see them going back to GCN-style design for UDNA.

Integrating matrix cores to RDNA4 compute unit would be an iterative step towards UDNA. They could iterate that compute unit further to meet AI workload.

I would expect UDNA to borrow lot of CDNA3 tech for caches, memory controllers and connectivity with chiplets.

There's rumors that RDNA4 will be a small iterative improvement as they are ditching RDNA arch one year later. But i'd say it's entirely possible that RDNA4 is a an iterative step towards UDNA. First time they merge new RDNA ALU pipes with CDNA matrix units. Could be a big step.

It's important to consider the importance of AI for AMD. RDNA architecture was initially designed for AMDs most important market segment (gaming), while UDNA got the old GCN design. Now AI is most important segment for them. They want newest arch for professional AI chips too.

One might argue that simple GCN-style scheduling is still good for AI workloads, but lower latency scheduling can operate better on register pressure. And it's often better for caches to avoid running too wide workload on every compute unit. RDNA/Nvidia style arch is just better.

UDNA makes perfect sense. AMD today wants matrix units for AI based upscaling (and other client AI workloads) in consumer space. And they want their latest compute unit architecture for AI chips. AI is now a priority. GCN-style arch is outdated. Nvidia/RDNA-style is better.

del42sa · Saturday at 12:23 AM

so basicaly:

CDNA4 = MI350
CDNA5/CDNA Next + RDNA5 = UDNA = MI400/new Radeon family

soresu · Sunday at 4:59 PM

del42sa said:
so basicaly:

CDNA4 = MI350
CDNA5/CDNA Next + RDNA5 = UDNA = MI400/new Radeon family

If the rumour is true and UDNA is coming in 2026 I'd say a domain specific (gfx focused) RDNA5 never existed in the first place (outside earlier conceptual or strategic roadmaps).

A 2026 release would mean it's probably been in the works for 2-3 years already.

blckgrffn · Sunday at 8:58 PM

soresu said:
If the rumour is true and UDNA is coming in 2026 I'd say a domain specific (gfx focused) RDNA5 never existed in the first place (outside earlier conceptual or strategic roadmaps).

A 2026 release would mean it's probably been in the works for 2-3 years already.

Which would make it even easier to believe they would rather focus on that than all the permutations of RDNA4. It’ll be a pivotal architecture.

soresu · Monday at 1:51 AM

blckgrffn said:
Which would make it even easier to believe they would rather focus on that than all the permutations of RDNA4. It’ll be a pivotal architecture.

I had the same thought myself.

Domain specific architectures are an impediment to their market ambitions as far as ecosystem simplification goes, so the faster they progress to UDNA the better.

That being said I wish they would hurry up and release RDNA4 already.

poke01 · Monday at 1:58 AM

blckgrffn said:
It’ll be a pivotal architecture.

In what way? I mean we heard the same with RDNA3 and look how that turned out. I really do hope UDNA is good because a GPU monopoly isn't good.

marees · Monday at 2:06 AM

soresu said:
I had the same thought myself.

Domain specific architectures are an impediment to their market ambitions as far as ecosystem simplification goes, so the faster they progress to UDNA the better.

That being said I wish they would hurry up and release RDNA4 already.

I think RDNA 4 being cut down to size was a combination of

UDNA ambitions
TSMC (packaging) constraints
RDNA 3 being a misfire

All of this still doesn't explain why we don't have a monolithic Navi 49 that replaces Navi 31, unless there is a UDNA Navi 53 waiting in the wings that can do it

soresu · Monday at 2:06 AM

poke01 said:
In what way?

In the way that ROCm while still very flawed has come a long way since the last time the AMD GPU ecosystem was unified with Vega.

soresu · Monday at 2:07 AM

marees said:
All of this still doesn't explain why we don't have a monolithic Navi 49 that replaces Navi 31

Simply that they have put all pedals to the metal on UDNA roadmap.

Discussion RDNA4 + CDNA3 Architectures Thread

Golden Member

Golden Member

Diamond Member

Platinum Member

Senior member

Senior member

Member

Lifer

Diamond Member

Senior member

Member

Senior member

Diamond Member

Member

Diamond Member

Senior member

Golden Member

Golden Member

Member

Diamond Member

Diamond Member

Diamond Member

Platinum Member

Senior member

Diamond Member

Diamond Member