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

 

[marees]

I wonder if it is a type and it should by Ryzen, rather than Radeon.

Also, there were some rumors that AMD is going to call Strix Halo line "Max". So perhaps the 4 memory channel platform will be called "Max" and there could be something lower end to plug into it than Strix Halo?

I too first thought it could be a typo.
But he put xtx there

He usually doesn't shit poast (but did one time, which has damaged his credibility forever, imo)

 

[RnR_au]

I wonder if it is a type and it should by Ryzen, rather than Radeon.

Has the Ryzen line ever used XTX ?

I think its a joke.

Edit: Or maybe an outside chance of a consumer based AI dedicated card with 32GB of vram... very outside...

 

[marees]

AMD aligning its ducks in a row ahead of RDNA 4 & 5 releases ??

Driver Experiments is a new tool that allows low-level control of the AMD Adrenalin driver. It is helpful for game developers and other graphics developers in debugging their applications if they crash or work incorrectly on an AMD GPU.

This tool exposes some of the driver settings that were previously only available to AMD engineers who develop the driver, e.g. disabling support for ray tracing or some optimizations in the shader compiler. It may be useful for debugging issues in graphics applications – alone or together with other tools, like Radeon GPU Detective (RGD), Radeon GPU Profiler (RGP), Radeon Memory Visualizer (RMV), or Radeon Raytracing Analyzer (RRA).

https://gpuopen.com/learn/rdts-driver-experiments/​

Available driver experiments​

• Features — to disable specific feature that impact performance​

  • Disable mesh shader support
  • Disable sampler feedback support
  • Disable raytracing support
  • Disable variable rate shading
  • Disable GPU work graphs support
  • Disable low precision support
  • Disable native 16-bit type support
  • Disable AMD vendor extensions
  • Disable compute queue support
  • Disable copy queue support

• Optimizations — disable the optimizations that are enabled by default​

  • Disable floating-point optimizations
  • Disable shader compiler optimizations
  • Disable barrier optimizations
  • Disable acceleration structure optimizations
  • Force shader wave size
  • Disable raytracing shader inlining
  • Disable shader cache

• Safety features — to debug type of error in the code which makes the application unstable​

  • Disable depth-stencil texture compression
  • Zero unbound descriptors
  • Thread-safe command buffer allocator
  • Force structured buffers as raw
  • Vertical synchronization

How to use it?​

To get started with the Driver Experiments feature, you need Radeon Developer Panel v3.2 or later. You will also need the following:

• GPU: AMD Radeon RX 5000, 6000, 7000 series card
• Operating system: Windows® 10, Windows® 11
• AMD Software: Adrenalin Edition: version at least 24.9.1

 

[soresu]

AMD aligning its ducks in a row ahead of RDNA 4 & 5 releases ??

Nah, just aligns with AMD's claimed new focus on software.

Drivers are pretty fundamental as GPU software goes.

 

[ToTTenTranz]

Yes, it was a joke.

https://twitter.com/x/status/1842046573343305894

 

[marees]

I think below article confuses RDNA 5/UDNA features (tensors/matrix multiplication) with RDNA 4 (double TOPS of RDNA 3 but same architecture)

As we get closer to RDNA 4 later this year or in 2025, it’s becoming more likely that AMD will wait to update FSR to use AI accelerators until the new hardware launches. After all, we know the dedicated upscaling technology used in the PS5 Pro is being co-engineered with AMD, and I would be shocked if Team Red didn’t work some of that magic juice into its upcoming graphics cards.

Nothing is really known about what the AMD Radeon RX 8000-series will look like, but there’s no way AMD doesn’t work some kind of next-generation AI accelerators into the GPU. The real trick will be to actually use them for gaming this time around. It’s time.

The PS5 Pro Proves AI Upscaling Is the Future

Looking into the future.

in.ign.com

 

[soresu]

I wish all these people like Everest would just move to BlueSky.

I'm not using Twitter.

 

[Mahboi]

I wish all these people like Everest would just move to BlueSky.

I'm not using Twitter.

What's the difference?

 

[soresu]

What's the difference?

I don't need to sign up to BlueSky to see a full timeline of posts.

If I go to Twitter without signing in the timeline is a complete unnavigable mess and has been for well over a year.

 

[marees]

RDNA:
It is rumored that AMD will launch the Radeon RX 8000 series graphics cards designed with RDNA 4 architecture in the first quarter of next year. The initial firepower will focus on the mid-to-low-end market to seek to expand more market share

CDNA ??:
The two major manufacturers have unanimously turned to TSMC to invest in wafers, making TSMC's advanced process orders continue to be full. Among them, Huida has ordered a large number of 4nm processes from TSMC, including HPC and GPUs for consumer and e-sports applications, which are currently in the mass production stage.

AMD is mass-producing new products using TSMC's 3nm wafers. The industry estimates that the momentum for wafer production next year is expected to increase quarter by quarter. We are optimistic about the sales momentum of AMD's new graphics cards, and board card manufacturers are also receiving customer orders one after another, and we are optimistic that they will enter the supply stage in the first half of next year.

It is understood that Huida and AMD's new graphics cards will enter the CoWoS packaging stage after being put into mass production by TSMC, and will be entrusted to ASE Silicon Products and the world's largest testing factory KYEC to conduct wafer testing (CP) and final test (FT), and at the same time, it is used with probe cards from test interface manufacturers such as Wangsi, Yingwei, and Jingce for testing, taking advantage of the trend to activate the entire semiconductor supply chain.

輝達、超微搶推新顯卡 半導體鏈備貨動起來 | 今晨必讀 | 要聞 | 經濟日報

輝達（NVIDIA）、超微（AMD）市場爭奪戰，從AI晶片一路打到繪圖處理器（GPU），兩強新一代AI晶片陸續面市之際，...

money-udn-com.translate.goog

 

[soresu]

CDNA ??:
The two major manufacturers have unanimously turned to TSMC to invest in wafers, making TSMC's advanced process orders continue to be full. Among them, Huida has ordered a large number of 4nm processes from TSMC, including HPC and GPUs for consumer and e-sports applications, which are currently in the mass production stage.

A recent AMD event had a slide showing that the post CDNA4 µArch in 2026 will be a next gen architecture, signifying the jump from GFX9 lineage finally.

 

[SolidQ]

post CDNA4 µArch in 2026

They seems going still use CDNA name, or they just split UDNA

 

[Kepler_L2]

They seems going still use CDNA name, or they just split UDNA

UDNA name is a recent invention. The internal name for the uarch is another star-related name.

 

[soresu]

UDNA name is a recent invention. The internal name for the uarch is another star-related name.

Betelgeuse.... Betelguese..... Betelgeuse.....

 

[soresu]

I seem to remember one of the CDNA µArchs being codenamed Aldebaran or something similar.

Ah yes, CDNA2.

CDNA1 is Arcturus.

CDNA3 is Antares.

 

[soresu]

CDNA4 could be Auriga or Alpha Centauri going by the A* naming convention of the GFX9+ lineage.

 

[SolidQ]

CDNA µArchs being codenamed Aldebaran

another star-related name.

That is next codename

 

[branch_suggestion]

CDNA4 could be Auriga or Alpha Centauri going by the A* naming convention of the GFX9+ lineage.

They've been in order of solar mass so far, so I'll guess Betelgeuse next.

 

[Glo.]

They've been in order of solar mass so far, so I'll guess Betelgeuse next.

And Betelgeuse is about to explode, soon.

 

[soresu]

And Betelgeuse is about to explode, soon.

Weeeeell my pedantic side can't help but comment that it probably went nova centuries ago already 😅

 

[SolidQ]

Interesting would AMD name N44 R8500? Like R200?

 

[Tuna-Fish]

I'd hope they'd name it 8500 for the full and 8400 for the cut down version, but I think that's not going to happen. The model number inflation has been going on for a long time now, and now even NV is selling a 128-bit bus card as the 60-series. Full N44 will likely be 8600.

 

[marees]

I am guessing 5070 = 4070 ti super ?

So n48xtx should be $100 cheaper ~ $500-$600 ??

 

[Saylick]

I am guessing 5070 = 4070 ti super ?

So n48xtx should be $100 cheaper ~ $500-$600 ??



View attachment 109055

JHH at CES: “Presenting billions of dollars of engineering time, the RTX 5090 with DLSS4 and RTX AI. For those who deserve the best, the Founder’s Edition retails for $2499.”

Me:

 

[Kepler_L2]

I am guessing 5070 = 4070 ti super ?

So n48xtx should be $100 cheaper ~ $500-$600 ??



View attachment 109055

If Jack is serious about getting 40-50% market share, he should launch N48 XTX for $399