Live data from Hacker News

CUDA on Windows Subsystem for Linux 2

devblogs.nvidia.com

51–60 of 221 posts

Re: CUDA on Windows Subsystem for Linux 2

#51

Throwing out a random data point: Until now, NVIDIA has been very anti-virtualization (on the consumer side), going as far as to engineer its drivers to detect in-use virtualization (Hyper-V, Xen, QEMU, etc.) and fail on purpose [1]. I'm curious to see how they now handle this scenario (given WSLv2 runs in a virtual machine). Perhaps they just commented those checks out in their 'specialized' drivers, an interesting…

It’s not GPU pass-through, it’s paravirtualization.

Re: CUDA on Windows Subsystem for Linux 2

#52
post #37
post #31

Earlier quoted context omitted.

Xming on Windows isn't exactly a walk in the park either -- I'd still rather use the native, proprietary build of VSCode on a proprietary OS with the SSH extension if I'm on a high-DPI screen.

https://x410.dev/ is trivial to set up, and once done, it "just works". But yeah, with VSCode WSL remoting, there's just no particular reason to do that anymore.

What's the performance like on this? Does it feel native or does it feel more like a VM?

Re: CUDA on Windows Subsystem for Linux 2

#53
post #46
post #31

Earlier quoted context omitted.

Xming on Windows isn't exactly a walk in the park either -- I'd still rather use the native, proprietary build of VSCode on a proprietary OS with the SSH extension if I'm on a high-DPI screen.

VSCode works just fine with hidpi on Linux. It’s just an Electron app after all, and Chromium has had hidpi support since forever.

Many many many apps on Linux still have massive issues with DPI, especially with mixed-DPI environments (which are no longer an edge case, they're the Common Case with a laptop attached to a monitor).

Even accessing machines remotely via Xrdp has huge issues because once you create the session with a certain DPI, logging into the session from a different DPI machine means you're stuck reading either extremely tiny or extremely huge text

Re: CUDA on Windows Subsystem for Linux 2

#54
post #37
post #31

Earlier quoted context omitted.

Xming on Windows isn't exactly a walk in the park either -- I'd still rather use the native, proprietary build of VSCode on a proprietary OS with the SSH extension if I'm on a high-DPI screen.

https://x410.dev/ is trivial to set up, and once done, it "just works". But yeah, with VSCode WSL remoting, there's just no particular reason to do that anymore.

I bought x410 but gave up later since I needed OpenGL3 among other things. It is nicely done, worth the 10$.

Re: CUDA on Windows Subsystem for Linux 2

#55
post #39

Earlier quoted context omitted.

Can you be more specific? What exactly is horrible on Windows when you've tried coding?

- Powershell sucks - Graphical installers / no package manager - No search results / docs for most languages - Weird bugs due to being second class for most languages - Have to use docker in VM - Bad env for writing scripts - We deploy on Linux

Some of these are fixed or there's progress being made:

- IMHO PowerShell is not that bad, but the newer PowerShell Core works better and with WSL you can use your favorite linux shell (and mix and match Windows and Linux binaries)

- WinGet is in preview

- No search results: my experience is 50/50 on that

- second class citizen: I wouldn't say most, definitely for Golang

- Docker now runs on WSL2 (which is still a sort of VM, but in my experience it's night and day with the old VM and especially compared to the default Docker experience on macOS)

- Why? You can use Python, and scripts are where Powershell really shines as a language

- That's the entire point of WSL I guess

Re: CUDA on Windows Subsystem for Linux 2

#56

Earlier quoted context omitted.

Well, CUDA is about the opposite of open source. It needs special hardware and a closed source driver. CUDA on Windows was already a big deal in engineering. Scientists and academics preferred Linux (and python, and docker) and Windows is playing catch up there to become compatible with the tools that have grown in popularity over the last 5 years. Personally, as an open source advocate, I think competition (even fro…

Also Linux is becoming the new Windows, and is one of the reason I stopped using it. What is the point of using a system that on every release is more and more bloated and full of crap that it breaks? Docker for example, I hate it, it serves no purpose beside wasting disk space and slowing down your machine. BTRFS is a bugged imitation of ZFS. Wayland not only is bugged but even does less things than Xorg. Most scien…

> Docker for example, I hate it, it serves no purpose beside wasting disk space and slowing down your machine.

Docker is a deployment tool and obviously needed by enterprises and I don't think anyone sane using it for their own localhost.

> BTRFS is a bugged imitation of ZFS.

You don't really have to use it though. ext works flawlessly.

> Wayland not only is bugged but even does less things than Xorg.

Check out this video from one of major X maintainer of the past:

https://www.youtube.com/watch?v=RIctzAQOe44

Re: CUDA on Windows Subsystem for Linux 2

#57

When you mention Microsoft + EEE, you get downvoted. Well, they have so far: 1) embraced Linux. 2) extended it with proprietary functionality (Windows living outside WSL), and now what is next? How is this not the same? Of course, the MS employee will downvote me automatically, the VS code users will downvote me too, and I will lose 5 karma for saying the truth.

Next? They're trying to upstream it

Re: CUDA on Windows Subsystem for Linux 2

#58
post #51

Throwing out a random data point: Until now, NVIDIA has been very anti-virtualization (on the consumer side), going as far as to engineer its drivers to detect in-use virtualization (Hyper-V, Xen, QEMU, etc.) and fail on purpose [1]. I'm curious to see how they now handle this scenario (given WSLv2 runs in a virtual machine). Perhaps they just commented those checks out in their 'specialized' drivers, an interesting…

It’s not GPU pass-through, it’s paravirtualization.

Still if they want they can allow only Quadro cards to expose the necessary interfaces on the host side.

Re: CUDA on Windows Subsystem for Linux 2

#59
post #57

When you mention Microsoft + EEE, you get downvoted. Well, they have so far: 1) embraced Linux. 2) extended it with proprietary functionality (Windows living outside WSL), and now what is next? How is this not the same? Of course, the MS employee will downvote me automatically, the VS code users will downvote me too, and I will lose 5 karma for saying the truth.

Next? They're trying to upstream it

Why would it matter if it's upstream? Upstreaming something like this is not really useful to anyone but Microsoft; it is still something that is inextricably tied to Windows and WSL.

The point is not "extinguishing" Linux per se, it's achieving enough lock-in that only Linux that Microsoft customers can use is WSL.

Re: CUDA on Windows Subsystem for Linux 2

#60
post #47

This is quite light on detail, have just installed, was thinking that nvcc would be installed to WSL2 as part of the process. It wasn't so did it via apt which worked fine. Built a sample (with some hacking of the SM versions which should be higher as I have an RTX) and when I run the basic Matrix mult demo get code=38(cudaErrorNoDevice) Anyone else had any luck?

Just done some more reading around and it seems most people are using docker in WSL to get things work! This seems overkill to me. I usually just write CUDA code in linux and use it. Am I missing something?
Post reply on HN