4 ms·
There is still no replacement for CUDA. ROCm is a compatibility joke, Vulkan is too insanely low-level, OpenCL runs way too slow. Compute shaders are technicall
by freeone3000 3y ago
There is still no replacement for CUDA. ROCm is a compatibility joke, Vulkan is too insanely low-level, OpenCL runs way too slow. Compute shaders are technically fine but bring flashbacks to 2006. Metal is a good alternative but only offered on systems without discrete GPUs! so we’re running CUDA for the next five years. In a conflict between Linux and Nvidia, in the HPC or AI spaces, Nvidia wins.
- tw04 3y ago>In a conflict between Linux and Nvidia, in the HPC or AI spaces, Nvidia wins. I HIGHLY doubt this. Nvidia has decades of investment in Linux internally. They aren't going to convert their DGX to Windows, or their dozens of internally utilized supercomputers. Sure, in a small window end-users are going to make a LOT of noise about Linux "breaking" - but unless Nvidia is planning on forking the kernel and fully maintaining their own version of Linux, they'll cave.
- MichaelZuo 3y ago> ... but unless Nvidia is planning on forking the kernel and fully maintaining their own version of Linux, they'll cave. Why wouldn't they? It would only be a few million dollars a year to do so, at most.
- seanw444 3y agoYeah, they totally would.
- COGlory 3y agoTheir customers are not going to run an Nvidia fork, because the vast majority of applications only compile on very specific Linux configurations to begin with. This would effectively cede the market to CPUs or AMD.
- MichaelZuo 3y agoNo, it would be a lot easier to modify their applications then wait for AMD to create CUDA 2.0.
- lamontcg 3y ago> but unless Nvidia is planning on forking the kernel and fully maintaining their own version of Linux https://docs.nvidia.com/dgx/dgx-os-6-user-guide/introduction.html#dgx-os-6-features https://docs.nvidia.com/dgx/dgx-os-6-user-guide/introduction...
- tw04 3y agoThat isn't a permanent fork away from the Linux kernel. It's basically them repackaging Ubuntu LTS, customized with some of the binaries packaged in. Permanently forking the Linux kernel us a VERY different beast then taking Ubuntu, pre-installing Nvidia drivers, and then re-baselining everytime Ubuntu releases a new LTS.
- lamontcg 3y agoEntirely depends on how extensive the patches are. FAANGS will all have their own kernels with custom patches, some of which will find their way back into the distros, and some which are kept in-house for custom reasons. It isn't necessarily that burdensome. And I was involved in doing this at Amazon (although a long time ago now). NVidia also doesn't really need to fundamentally fork the linux kernel to do this since drivers are already a pretty hard API boundary layer. They won't need to touch the TCP/IP stack, scheduler, etc. Their version of a "fork" would be "add back the API which is being removed here".
- nvm0n2 3y agoLol, no they won't. The cost of developing their software stack is massive. They'd either resort to more technical hacks like running more of the driver in userspace, or if push comes to shove just develop a Linux compatible kernel with a cleanroom codebase. Kernels aren't that hard to develop for a company like NVIDIA especially if you only care about supporting a limited set of hardware (like DGX servers).
- lucasyvas 3y agoLinux wins hands down, it always has.
- elromulous 3y ago+1 Nvidia would win this fight. Linus and co taking a hard stance would result in the industry bifurcating on Linux. AI just has so much momentum and money being poured into it.
- sam0x17 3y agowebgpu will eventually fill this role right?