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

 

[basix]

- CDNA5 is rumored to use same CU and cache structure as RDNA5 (altough different sizes)
- Universal Compression could be a transparent drop-in IP around caches and memory subsystem. Should not require a major rework of the SoC and memory architecture
- Neural Arrays would be a natural fit for ML accelerators

If I had to guess, CDNA5 will share many similarities to RDNA5.
gfx1250 indicates at least a spin-off from RDNA4. But as RDNA5 will arrive some time after CDNA5, gfx1250 for CDNA5 and gfx1300 for RDNA5 somewhat makes sense.

 

[MrMPFR]

I've replied to @basix's comment on the RDNA 5 thread. Can we please continue any Nextgen GPU discussions there and not here xD

 

[SolidQ]

Microsoft Update Catalog

Windows 11 KB5072033 for 25H2 and 24H2

• Gaming performance has been enhanced, particularly for AMD GPU users, with fixes for issues like "GPU hung" and "driver removed" errors, and resolution of a bug that incorrectly flagged supported graphics cards as unsupported.

P.S Ordered for tomorrow Powercolor Hellhound RX 9070XT, will test after new year.

 

[Arielrt17]

So Both, Radeon AI Pro 9700s and 9600D are Fanless Cards for Edge AI, looks like based on the same Navi 48 Die.

https://www.amd.com/en/products/gra.../ai-9000-series/amd-radeon-ai-pro-r9700s.html

https://www.amd.com/en/products/gra.../ai-9000-series/amd-radeon-ai-pro-r9600d.html

 

[maddie]

So Both, Radeon AI Pro 9700s and 9600D are Fanless Cards for Edge AI, looks like based on the same Navi 48 Die.

https://www.amd.com/en/products/gra.../ai-9000-series/amd-radeon-ai-pro-r9700s.html

https://www.amd.com/en/products/gra.../ai-9000-series/amd-radeon-ai-pro-r9600d.html

300W and no cooling fan, how does that work? System fans force air lengthwise through fins.

Or, render artist is nontechnical?

 

[poke01]

It’s been a couple of weeks with my 9070xt card and I love it.

No crashes, no timeouts. I set the limit to 243w and -65v. I lose like 10 fps and 4000 points in time spy but it’s a much quieter card and doesn’t go over 90c for the hotspot.

 

[marees]

The first rdna 4 apu ???

Exynos 2600 has 8 wgp rdna 4 apu ??

https://twitter.com/x/status/2004449054756163769

 

[marees]

The first rdna 4 apu ???

Details

4.4. RDNA 4 based series​

4.4.1. MGFX4 (gc_40_4)​

It is a first-generation RDNA 4-based custom GPU architecture developed by Samsung Electronics. The first appearance is Xclipse 960 mounted on Exynos 2600. Computation performance has doubled compared to Xclipse 950, and when ray tracing is activated, FPS increases by 50% compared to the previous generation. Additionally, an upscaling and frame interpolation technology called ENSS (Exynos Neural Super Sampling) was introduced for the first time.

The code name is Juno .

List of processors used
mounted AP (part number)	GPU name	GPU	ANGLE version (Vulkan)	system memory
SP:TU:ROP (GP, PU, RA, RZ)	clock (MHz)	L2 cache (MB)	bus (Bit) (channel)	standard	clock (Bitrate) (MHz) (Mbps)	allocated capacity (GB)
Exynos 2600 (S5E9965)	Xclipse 960	1024:64:64 (1, 2, 16, 2)	980	2	1.4.304	(16) (4)	LPDDR5X	5333 (10667)	4

https://en.namu.wiki/w/%EC%82%BC%EC%84%B1%EC%A0%84%EC%9E%90/%EB%A7%88%EC%9D%B4%ED%81%AC%EB%A1%9C%EC%95%84%ED%82%A4%ED%85%8D%EC%B2%98#s-4.4

 

[NTMBK]

The first rdna 4 apu ???

Sure would be nice if that post actually cited any verifiable sources 🤨

 

[Win2012R2]

So it appears Samsung licensed RDNA4, they will certainly not want to manufacture it on TSMC, which means AMD must have Samsung compatible RDNA4 stuff, maybe they will use them for RDNA5...

 

[marees]

So it appears Samsung licensed RDNA4, they will certainly not want to manufacture it on TSMC, which means AMD must have Samsung compatible RDNA4 stuff, maybe they will use them for RDNA5...

I am expecting something like zen 7 — the lowest powered sku to replace mendocino — grimlock point 4 with RDNA 5

 

[SolidQ]

RX 9070 appear in steam

They need split data for 7800XT and 9070XT, i assume most of them is RX 9070XT

0.69% for one month is too much for 7800XT

 

[MrMPFR]

RX 9070 appear in steam

They need split data for 7800XT and 9070XT, i assume most of them is RX 9070XT

0.69% for one month is too much for 7800XT

Entire dataset is broken. DX12 = 131.44% and DX8 = -32.39% xD

 

[SolidQ]

Wonder, if AMD have answer to DLSS 4.5 and MFG 6. Maybe FSR 4.5 with transformer mode and at least Mfg 4x

 

[adroc_thurston]

Wonder, if AMD have answer to DLSS 4.5 and MFG 6. Maybe FSR 4.5 with transformer mode and at least Mfg 4x

No they don't.

 

[MrMPFR]

• No MFG anywhere in AMD pipeline - 2027 or later
• Abhorrent frame pacing - no update = 2027 or later
• No joint SR and RR - as per FSR blog still in R&D phase, paper here, no firm timeline so 2027 or later
• Inferior motion clarity to DLSS4, closer to DLSS 3.7 - maybe they can get to DLSS4 SR fidelity by EoY 2026

Now NVIDIA is moving goalpost yet again. Prob significantly.

Oh and @SolidQ FSR4 is already hybrid ViT+CNN. AMD just put this together hastily last minute + still higher perf overhead than DLSS4.
By the time AMD has what you're asking for NVIDIA is prob already at DLSS5 with Ultra models reserved for 60 series. IIRC CPX extrapolation was 8x raw NVFP4 dense vs 50 series. IF DLSS (currently FP8) works with NVFP4 then that's 16X matmul at iso-core + clocks.

Good luck AMD and thank you David Wang

 

[SolidQ]

No MFG anywhere in AMD pipeline - 2027 or later

I assume full transformer+ and mfg, they left for RDNA5 matrix cores + new things.

Anyway still can't find games, where i need upscaling. Just bought Horizon Forbidden West to test RX 9070XT.
Still deciding get super cheap B850 for 9700x, and then B950 for zen7, or X870 tomahawk, and then upgrade to Zen7

 

[Kepler_L2]

Wonder, if AMD have answer to DLSS 4.5 and MFG 6. Maybe FSR 4.5 with transformer mode and at least Mfg 4x

No one cares about MFG

 

[Tup3x]

No one cares about MFG

Works surprisingly well especially when CPU would be bottlenecking. I'd take 60 fps with 3x frame gen over native any day (in single player games anyway). In fact, if game is stuck with 60 fps limit I enabled driver frame gen, because it just doesn't look smooth.

 

[SolidQ]

No one cares about MFG

Yes, but working as marketing

 

[Mopetar]

Works surprisingly well especially when CPU would be bottlenecking. I'd take 60 fps with 3x frame gen over native any day (in single player games anyway). In fact, if game is stuck with 60 fps limit I enabled driver frame gen, because it just doesn't look smooth.

Frame gen is only any good when you already have a high enough FPS not to need it. For anything less it just adds latency and there's little point in getting 60 FPS if the game is still only receiving and reacting to input at a third that rate. It might look smoother, but it still controls like driving through mud.

But it makes for bigger bars. I'll bet whatever JHH and crew are cooking up next will make the bars shiny. You wouldn't want to buy a GPU that doesn't have shiny bars in the benchmark charts would you?

 

[T2098]

300W and no cooling fan, how does that work? System fans force air lengthwise through fins.

Or, render artist is nontechnical?

Rackmounted servers are normally an insanely noisy windtunnel of airflow from front to back, and data center focused products commonly rely on that airflow, which is usually on the product spec sheet somewhere.
It's common to take datacenter 10Gb/25GB/40GBe cards and put them into a consumer desktop PC and have them overheat nearly immediately, as their passive heatsinks were designed with that very high amount of overall system airflow to stay in spec.

Most datacenter-specific cards are built like that for maximum density.
The sort of systems those are meant to go in are generally built like this:

 

[basix]

No one cares about MFG

If you are already using FG (which I do), MFG brings additional motion smoothness without much additional latency. So I would probably use it if my GPU would support it. It will not be a game changer but still nice to have.

 

[blackangus]

If you are already using FG (which I do), MFG brings additional motion smoothness without much additional latency. So I would probably use it if my GPU would support it. It will not be a game changer but still nice to have.

I have to take a moment here to agree.
I was the biggest - No Frame Gen BS! - Guy.

However this last PC build I had to compromise to get a video card and ended up with a 5080 and a 240hz 4k OLED.
This combo of frame-gen and 240 hz is really what frame gen is about.

What it is - Generate frames on top of already good FPS (60 FPS +) to get buttery smoothness from your HFR monitor. (Assuming you are not already at max FPS)
What it is NOT - A mechanism to make low FPS playable (yes it helps but lag is pretty bad)

I would literally never go back to non-frame gen unless my card could do the work natively.

 

[ToTTenTranz]

What it is NOT - A mechanism to make low FPS playable (yes it helps but lag is pretty bad)

Framegen on top of low FPS also creates more discernible artifacts, not to mention the fact that it doesn't work very well with adaptive sync so stuttering gets more noticeable.
I have a 165Hz monitor but I don't really use framegen even on the RTX5090. The setbacks are a bit too great for me, IMO. I prefer cranking up IQ until I reach 80-90FPS.