Updated NVIDIA CUDA For WSL Brings Better Performance, PTX JIT

[phoronix]

Phoronix: Updated NVIDIA CUDA For WSL Brings Better Performance, PTX JIT

Earlier this summer building off the latest Windows Subsystem for Linux 2 advancements by Microsoft, NVIDIA released early support for CUDA / GPU compute on WSL2. This week NVIDIA offered up a new version of their CUDA WSL support...

http://www.phoronix.com/scan.php?pag...UDA-WSL-Update

[evasb]

Embrace: WSL2

Extend: https://www.phoronix.com/scan.php?pa...soft-DX12-WSL2

...and now we entered in the Extinguish phase.

[pipe13]

Embrace: WSL2

Extend: https://www.phoronix.com/scan.php?pa...soft-DX12-WSL2

...and now we entered in the Extinguish phase.

No one says you have to use Nvidia. No one says you have to use Windows. There are viable alternatives to both.

[JPFSanders]

Amusing, Nvidia is just amusing. At this point their refusal to either allow the community to create a working FOSS driver, or they themselves providing with one feels like outright purposeful sabotage.

[evasb]

No one says you have to use Nvidia. No one says you have to use Windows. There are viable alternatives to both.

nVidia has the GPGPU market locked-in with CUDA. If nVidia just drops Linux, or just launch barebones drivers for real Linux, in favour of WSL we are done.

[kpedersen]

No one says you have to use Nvidia. No one says you have to use Windows. There are viable alternatives to both.

I have worked for companies where both of those statements would be contended.

All upper management knows the *only* correct way is Nvidia and Microsoft.

(Even if they are both scum and should be avoided at all costs)

[zexelon]

No one says you have to use Nvidia. No one says you have to use Windows. There are viable alternatives to both.

Unfortunately this is not true. There are no alternatives to CUDA yet. If you are building a production ML platform depending on the maximum level of support for your customers there is nothing else. AMD has had something like three major false starts on platforms for GPGPU and this has really made them look like a joke in this arena. AMD GPU hardware also just cant compete with Nvidia at this point, especially with the Ampere platform out now. I cant wait to get my hands on an A100 server some day!

To echo what evasb said, if Nvidia drops Linux for whatever reason, Linux will be regressed back to the 1990s user base.

[mcoffin]

For someone with as much invested in the Linux and free software ecosystem as you, I wish you would have had to forethought not to drum up hype around this transparent embrace, extend, extinguish attempt.

Keep it up. I love your content, but adding fuel to the hype train Microsoft is building for WSL borders on naivety. I "can't wait" until we start seeing "GNU/Linux" programs that are built only for WSL.

Don't get me wrong; this is newsworthy, but there's something to be said for being complicit in this whole happening. I don't think you'll want to have been a part of it when this has all played out.

[Paradigm Shifter]

Unfortunately this is not true. There are no alternatives to CUDA yet. If you are building a production ML platform depending on the maximum level of support for your customers there is nothing else. AMD has had something like three major false starts on platforms for GPGPU and this has really made them look like a joke in this arena. AMD GPU hardware also just cant compete with Nvidia at this point, especially with the Ampere platform out now. I cant wait to get my hands on an A100 server some day!

To echo what evasb said, if Nvidia drops Linux for whatever reason, Linux will be regressed back to the 1990s user base.

Any sort of production pseudo-HPC workstation, not just ML.

As much as I'd like to disagree, I really can't. I haven't used a desktop AMD GPU since the 4870X2 (although to be fair I did briefly toy with with 6970 which was given to me) solely because of CUDA. The mess that was HSA, and the mess that is ROCm just makes me despair. Intel stand more chance of getting oneAPI fully off the blocks and running as a functional competitor to CUDA (with the Xe GPUs and AVX-512, as much as I dislike the latter) before ROCm reaches a point where I can get it working - reliably - across a wide range of systems without having to actually think beyond "what compute model is this GPU?"

What I really want to be able to do with ROCm is have it working without workarounds, bodges or hacks on an APU, because justifying buying an APU-based system is much easier to either a) myself or b) my boss, when all it will really be is a "tinkering/testing" rig. Instead, I basically have to get a Radeon VII (or two) which is an impossible sell without serious evidence it will do exactly what we want.

I'm not sure nVidia will drop Linux - certainly not in favour of Windows - although as I said in a previous thread, with them after ARM, and having already acquired Mellanox, I can see nVidia pulling an Apple and in-housing a *BSD clone to keep everything in one place. It wouldn't be too horrific a shift for most academic/industrial users to move from Linux to BSD, although there will always be those who complain.

[zexelon]

Paradigm Shifter I like your idea there of a dedicated BSD built for their hardware. If they 100% switched over to say an in house clone of FreeBSD I think that would be a very interesting move! Certainly a heck of a better move then going windows only! When was there ever a top500 HPC system running windows :P