Re: Quick Poll: Debian to better support hardware acceleration?
Hello, I'm not confortable in fostering proprietary solution like CUDA against libre alternative in Debian project. CUDA libraries are de-facto outdated when new Debian release comes out due to new hardware release and vendor lock-in business model. As a consequence, the user will always download the last libraries version from the vendor web-site. So, spending time pushing these libraries in the archive is pointless in my opinion. I would better spend time promoting free alternative or high level abstractions like OpenCL or SYSCL even if they are less performant. One day they will be better and Debian would be part of the success. Reminds me some Direct3D/OpenGL war some times ago. With this in mind, I also think that Debian should not prevent CUDA integration in scientific softwares, maybe by providing a simple way to rebuild or configure software to use CUDA libraries from the nVidia website. Sorry for being more purist than pragmatic when dealing with Debian. As a reminder: https://www.debian.org/intro/why_debian.en.html Best, François Le vendredi 21 mai 2021 à 04:40 +, M. Zhou a écrit : > Hi folks, > > --- > > Q: How far should Debian go along the way for supporting hardware > acceleration solutions like CUDA? > > Choice 1: this game belongs to the big companies. we should offload > such burden to third-party providers such as Anaconda. > Choice 2: we may try to provide what the users need. > Choice 3: > > --- > > As we know, hardware acceleration means a lot to scientific > computing, > and I believe a number of debian users use solutions like CUDA, ROCm, > or even SYCL. And the most prevalent solution seems to be CUDA. > Recall that anaconda might be one of the simplest ways to get the > cuda > version of tensorflow and pytorch, etc. So I just want to hear your > opinions on how far we should go along this direction. > > If we really want to go further, then a GPU server should be > available > in our infrastructure to facilitate development. Although license is > another considerable blocker, this can be discussed later. > > Thanks! > signature.asc Description: This is a digitally signed message part
Re: Quick Poll: Debian to better support hardware acceleration?
On 21.05.21 21:13, François Mazen wrote: > CUDA libraries are de-facto outdated when new Debian release comes > out due to new hardware release and vendor lock-in business model. As > a consequence, the user will always download the last libraries > version from the vendor web-site. That's not entirely accurate, though. The Debian NVIDIA Team does a spectacular job of maintaining up-to-date backports in the official archive. I imagine that like myself, many other users will prefer the backport to the vendor version, as the backport certainly integrates more nicely. Best, Christian
Re: Quick Poll: Debian to better support hardware acceleration?
Am 21.05.21 um 15:55 schrieb Thomas Schiex: > I'm a computer scientist working in AI and structural biology. I'm > sorry to say that CUDA has slowly invaded a lot of our scientific > pipelines, for Deep learning, convex optimization and molecular > simulations. > > I just could not vote for option 2 even if option 1 is tolerable (I'm > using it). > > Le 21/05/2021 à 15:35, Julien Puydt a écrit : >> Le vendredi 21 mai 2021 à 04:40 +, M. Zhou a écrit : >>> Q: How far should Debian go along the way for supporting hardware >>> acceleration solutions like CUDA? >>> >>> Choice 1: this game belongs to the big companies. we should offload >>> such burden to third-party providers such as Anaconda. Once you decide to go the conda/brew/guix route, you are likely to stick to it also for your applications. Debian is then left out. >>> Choice 2: we may try to provide what the users need. Please. Otherwise we fail them. >>> Choice 3: >> I'm not a user of anything like it (as far as I know...), but it's >> Debian's mission to make useful software available : choice 2. Free CUDA drivers will take a bit longer to surface. We should have the non-free now and the free shall be supported - happily so with Debian Money if that changes anything. But our users should not need to wait. ROCm is late in the game, which is unfortunate. Would be great to have also more from Xilinx and Altera/Intel in our distribution to detect and program their FPGA. Thanks! Steffen
Re: Quick Poll: Debian to better support hardware acceleration?
I've had a massive faff with ROCm, and on my main workstation have been compelled to use a different distro that makes things easier. Definite preference for option 2. Previous job was academic research, everything was CUDA. Now I'm in industrial R, and I run into whatever my customers want me to use, which has been a mixture of CUDA and ROCm. With some practice and tinkering, CUDA can be made to play nicely. Sam On Fri, 21 May 2021 at 15:27, Thomas Schiex wrote: > I'm a computer scientist working in AI and structural biology. I'm sorry > to say that CUDA has slowly invaded a lot of our scientific pipelines, > for Deep learning, convex optimization and molecular simulations. > > I just could not vote for option 2 even if option 1 is tolerable (I'm > using it). > > Le 21/05/2021 à 15:35, Julien Puydt a écrit : > > Le vendredi 21 mai 2021 à 04:40 +, M. Zhou a écrit : > >> Q: How far should Debian go along the way for supporting hardware > >> acceleration solutions like CUDA? > >> > >> Choice 1: this game belongs to the big companies. we should offload > >> such burden to third-party providers such as Anaconda. > >> Choice 2: we may try to provide what the users need. > >> Choice 3: > > I'm not a user of anything like it (as far as I know...), but it's > > Debian's mission to make useful software available : choice 2. > > > > JP > > > >
Re: Quick Poll: Debian to better support hardware acceleration?
Hi, "M. Zhou" writes: > Q: How far should Debian go along the way for supporting hardware > acceleration solutions like CUDA? I think Debian is already doing a good job with CUDA, at least as it pertains to my work with Python+GPU. My thanks and please keep it up! One recent example for me is installing "cupy" from PyPI via pip. On Debian testing/sid it Just Works(tm)! On Ubuntu 18.04 it caused me to enter "CUDA install hell". The CUDA version provided by the OS is apparently not supported by "cupy" on PyPI. My attempts to build cupy from source against Ubuntu CUDA or to install more recent CUDA from non-Ubuntu to satisfy PyPI's cupy both were utter failures. For this and other reasons, my solution is actually to migrate my Ubuntu 18.04 systems to Debian. BTW, there is a HN thread formed this morning on Python/GPU things that may be usefully related here: https://news.ycombinator.com/item?id=27205586 https://discuss.python.org/t/what-to-do-about-gpus-and-the-built-distributions-that-support-them/7125 -Brett. signature.asc Description: PGP signature
Re: Quick Poll: Debian to better support hardware acceleration?
I'm a computer scientist working in AI and structural biology. I'm sorry to say that CUDA has slowly invaded a lot of our scientific pipelines, for Deep learning, convex optimization and molecular simulations. I just could not vote for option 2 even if option 1 is tolerable (I'm using it). Le 21/05/2021 à 15:35, Julien Puydt a écrit : Le vendredi 21 mai 2021 à 04:40 +, M. Zhou a écrit : Q: How far should Debian go along the way for supporting hardware acceleration solutions like CUDA? Choice 1: this game belongs to the big companies. we should offload such burden to third-party providers such as Anaconda. Choice 2: we may try to provide what the users need. Choice 3: I'm not a user of anything like it (as far as I know...), but it's Debian's mission to make useful software available : choice 2. JP
Re: Quick Poll: Debian to better support hardware acceleration?
Le vendredi 21 mai 2021 à 04:40 +, M. Zhou a écrit : > > Q: How far should Debian go along the way for supporting hardware > acceleration solutions like CUDA? > > Choice 1: this game belongs to the big companies. we should offload > such burden to third-party providers such as Anaconda. > Choice 2: we may try to provide what the users need. > Choice 3: I'm not a user of anything like it (as far as I know...), but it's Debian's mission to make useful software available : choice 2. JP
Re: Quick Poll: Debian to better support hardware acceleration?
On 21.05.21 06:40, M. Zhou wrote: > Choice 1: this game belongs to the big companies. we should offload > such burden to third-party providers such as Anaconda. > Choice 2: we may try to provide what the users need. > Choice 3: Choice 2, by a mile. CUDA wins either way. It's the de facto standard. It is also usable with commodity hardware available to and affordable by regular users. I think Debian loses either way, but not supporting it is the worse of the two options. Being purist here doesn't affect CUDA at all (see above), it just means Debian users will use pip/conda to get the software they need. Or worse, switch to a platform where CUDA is supported out of the box. And no acceleration is basically a non-option for certain workloads today. Nobody is going to do deep learning on a CPU. I'd strongly prefer ROCm as an acceleration solution (for obvious reasons), but the software environment is still far behind CUDA, and the officially supported hardware is unaffordable to the average user. We can work on the former, and hope for change on the latter, but until both of these issues are solved, we'll just lose people to pip/conda.
Re: Quick Poll: Debian to better support hardware acceleration?
CUDA is a poor solution. nVidia only, proprietary. Need to enourage upstreams to use the other solutions. I don't know ROCm, maybe it's good. Open source at least. SYCL is a good target, should be supported. Supersedes OpenCL. Drew On 2021-05-21 06:40, M. Zhou wrote: Hi folks, --- Q: How far should Debian go along the way for supporting hardware acceleration solutions like CUDA? Choice 1: this game belongs to the big companies. we should offload such burden to third-party providers such as Anaconda. Choice 2: we may try to provide what the users need. Choice 3: --- As we know, hardware acceleration means a lot to scientific computing, and I believe a number of debian users use solutions like CUDA, ROCm, or even SYCL. And the most prevalent solution seems to be CUDA. Recall that anaconda might be one of the simplest ways to get the cuda version of tensorflow and pytorch, etc. So I just want to hear your opinions on how far we should go along this direction. If we really want to go further, then a GPU server should be available in our infrastructure to facilitate development. Although license is another considerable blocker, this can be discussed later. Thanks!