WingNews logo WingNews
top | item 24465809

(no title)

nolaspring | 5 years ago

I spent most of this afternoon tying to get cuda in docker to work on my Mac for a machine learning use case. It doesn’t. Because nvidia

discuss

order

diesal11|5 years ago

"Because NVIDIA" is blatantly false.

CUDA support for docker containers is provided through the open source Nvidia-Docker project maintained by Nvidia[1]. If anything this is a great argument for NVIDIAs usage of open source.

Searching that project's issues shows that Nvidia-Docker support on MacOS is blocked by the VM used by Docker for Mac(xhyve) not supporting PCI passthrough, which is required for any containers to use host GPU resources.[2]

xhyve has an issue for PCI passthrough, updated a few months ago, which notes that the APIs provided by Apple through DriverKit are insufficient for this use case[3]

So your comment should really say "Because Apple"

[1] https://github.com/NVIDIA/nvidia-docker

[2] https://github.com/NVIDIA/nvidia-docker/issues/101#issuecomm...

[3] https://github.com/machyve/xhyve/issues/108#issuecomment-616...

kllrnohj|5 years ago

Apple is why you can't use Nvidia hardware on a Mac, not Nvidia. Apple has exclusive control over the drivers. Nvidia can't release updates or fix things on Mac OS.

I'm all for railing on the shitty things Nvidia does do, but no reason to add some made up ones onto the pile.

jefft255|5 years ago

« I spent most of the afternoon hacking away at some unsupported edge case on an unsupported platform which is inadequate for what I’m trying to do. It doesn’t work, which is clearly nvidia’s fault. »

fomine3|5 years ago

Because Apple. NVidia was tried to support Macs but Apple stops support.