News AMD unveils ROCm 7 — new platform boosts AI performance up to 3.5x, adds Radeon GPU support

[Admin]

AMD launches ROCm 7 platform that boosts AI inference performance by up to 3.5X, adds support for FP4/FP6 formats and distributed workloads, and extends compatibility to Windows and Radeon GPUs on Ryzen systems.

AMD unveils ROCm 7 — new platform boosts AI performance up to 3.5x, adds Radeon GPU support : Read more

 

[virgult]

...yet there's no mention of which Radeon GPUs are supported. There's conflicting information on the website. Here this "hardware spec" list seems to imply that the stack supports most modern Radeons down to the 6600 (plus Vega, the only arch ever really supported by ROCm anyway):
https://rocm.docs.amd.com/en/latest/reference/gpu-arch-specs.html

But this page only mentions Radeon Pro W 7000 as the cards one should buy to run ROCm 7:
https://www.amd.com/en/developer/resources/rocm-hub/dev-ai.html#tabs-d13ff37f6d-item-94bf1a0f36-tab

Let's see if AMD has chosen to be competitive with Nvidia from 2020 or still prefers to be stuck in 2017.

 

[bit_user]

...yet there's no mention of which Radeon GPUs are supported.

The place to look will be in the release notes. I don't yet see them for release 7, but here's the place in the 6.4.1 release notes, where hardware compatibility is specified:

https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html
​

 

[abufrejoval]

"The biggest change brought by ROCm 7 for client PCs is the extension of ROCm to Windows and Radeon GPUs, which allows the use of discrete and integrated GPUs for AI workloads, but only on Ryzen-based PCs."

I wonder where that comes from and if it will in fact be true. There is no technical reason why you shouldn't be able to run ROCm 7 workloads on an RX 9000 dGPU using an Intel CPU underneath.

It might not be able to use whatever Intel xPU facilities reside within Intel's APUs or SoCs, but for discrete GPUs that would be emulating some of Intel's worst behavior.

But what litlte information is out there only gives "positive" discrimination in that it promises support for the latest "AI" APUs and RX9000 series GPUs.

I doesn't mention RDNA 3, so unless they are supported already consumer type APUs and GPUs of these older generations seem likely to miss out.

In any case this is only an announcement so far, nothing I'd base a purchasing decision on before it's been validated by the likes of Phoronix.

 

[ejolson]

I think "but only Ryzen-based PCs" means the integrated GPUs on Ryzen processors are the only ones that work with ROCm 7. The need to say such an obvious thing makes more sense when one considers that HIP is cross-platform with backends to target both Nvidia and AMD accelerators.

Note also AMD had a line of APUs with integrated Radeon graphics before Ryzen that could to some extent be used for GPGPU computing. These will not work with ROCm.

 

[bit_user]

"The biggest change brought by ROCm 7 for client PCs is the extension of ROCm to Windows and Radeon GPUs, which allows the use of discrete and integrated GPUs for AI workloads, but only on Ryzen-based PCs."

I wonder where that comes from and if it will in fact be true.

Perhaps the "only on Ryzen-based PCs" was meant as a qualification only applying to the integrated GPUs?

 

[bit_user]

The need to say such an obvious thing makes more sense when one considers that HIP is cross-platform with backends to target both Nvidia and AMD accelerators.

HIP is designed as a cross-platform solution, but I have no idea about the current state of its support for running atop Intel iGPUs.

I consider HIP a vendor-specific API, even though the theoretical (and maybe rather real) possibility exists to use it atop other hardware. We cannot depend on them to maintain other backends with the same attention to quality, reliability, accuracy, and functional completeness as they do for their own hardware.

If I truly wanted a hardware-agnostic solution, I would opt for something that's not backed by a specific vendor, because they're always going to be biased towards their own hardware.