News Russian firm starts shipments of HPC system based on homegrown CPU — but cannot really avoid using a foreign GPU

[Admin]

Russian company Graviton launches first AI and HPC machine with domestic CPUs that will have to be used with Nvidia's H100.

Russian firm starts shipments of HPC system based on homegrown CPU — but cannot really avoid using a foreign GPU : Read more

 

[bit_user]

The Graviton S2124B server ...

The Graviton name is unfortunate, since they have nothing to do with Amazon's Graviton CPUs. I wonder which used that name first. Of course, the name itself dates back to almost 100 yeas ago, referring to a hypothesized elementary particle in the theory of quantum gravity.

those who use the S2124B will have to rely on Nvidia's CUDA ecosystem. However, without support from Nvidia, it is unlikely that peak performance will be achieved.

Nvidia has been developing an open source kernel driver. I haven't been following developments very closely, but a plausible scenario is that they use that and Rusticle to support running OpenCL on them. I'm not sure how good Rusticle is for running larger and more complex OpenCL apps, but progress on it has seemed fairly brisk.

Anyway, there's a scenario for you on how they could theoretically deliver this in usable form, even without CUDA. Performance would surely suffer, but it should (eventually) be usable for most purposes, I think.

Realistically, I'd expect they're going to find some way to get CUDA installed and working on it. CUDA does support the host platform running ARM CPUs, of course.

 

[Joseph_138]

The Graviton name is unfortunate, since they have nothing to do with Amazon's Graviton CPUs. I wonder which used that name first. Of course, the name itself dates back to almost 100 yeas ago, referring to a hypothesized elementary particle in the theory of quantum gravity.


Nvidia has been developing an open source kernel driver. I haven't been following developments very closely, but a plausible scenario is that they use that and Rusticle to support running OpenCL on them. I'm not sure how good Rusticle is for running larger and more complex OpenCL apps, but progress on it has seemed fairly brisk.

Anyway, there's a scenario for you on how they could theoretically deliver this in usable form, even without CUDA. Performance would surely suffer, but it should (eventually) be usable for most purposes, I think.

Realistically, I'd expect they're going to find some way to get CUDA installed and working on it. CUDA does support the host platform running ARM CPUs, of course.

I doubt that Nvidia is going to relinquish control of their drivers to the community. They've been very tightfisted about that in the past. That's why you could only ever get third party drivers like the Omega drivers, and Nimez drivers for AMD cards. Nvidia would not release the specifications and development tools to third parties, AMD would.

 

[bit_user]

I doubt that Nvidia is going to relinquish control of their drivers to the community. They've been very tightfisted about that in the past.

It's a done deal. They went public about it all the way back in May 2022.

NVIDIA Transitions Fully Towards Open-Source GPU Kernel Modules | NVIDIA Technical Blog

With the R515 driver, NVIDIA released a set of Linux GPU kernel modules in May 2022 as open source with dual GPL and MIT licensing. The initial release targeted datacenter compute GPUs…

developer.nvidia.com

However, the kernel driver is only one piece of their stack. A lot of their "special sauce", like CUDA, happens in userspace libraries. That's why I said you could theoretically cobble together a non-CUDA solution involving their hardware, because they have not open-sourced CUDA.

Nvidia would not release the specifications and development tools to third parties,

Right, and I still don't think they did. The open source driver is developed in-house.

AMD would.

They do publish some stuff, but it's mostly documentation of their shader ISA and I think not the full set of information that you'd need to write your own driver, if anyone would want to do that. AMD and Intel both have open source drivers for their GPUs as their only supported Linux option and they were both developed in-house (I think with some help from contractors like Redhat and Collabora; Valve also helped AMD).

AMD still maintains a set of proprietary userspace components, but that's mainly targeted at workstation users. For basically everyone else, they can just use Mesa and have a 100% open source stack for AMD.

I think Intel has had a 100% open source stack for quite a while, and it's their only supported option.