Ask HN: How to build GPU compute marketplace?

1 points by michaeloblak 3y ago ↗ HN

Is it possible? Let's say Alice has 2 GPUs idle at the moment, Ben has 1 GPU, and Chris needs 3 GPUs for the next 12 hours. How to build such a system, and what problems there might occur? How to handle turning off one's machine? Does it even make sense to run training or inference on such (latency)?

4 comments

[ 3.1 ms ] story [ 23.4 ms ] thread

ftxbro 3y ago ↗

There is already some like this including vast dot ai.

michaeloblak 3y ago ↗

Yeah, but how does it work exactly? Dropbox is rsync with UI, Instagram is a blog with pictures, but this?

How to access one’s compute? Overriding GPU drivers?

deserialized 3y ago ↗

It's just a remote docker context on a machine with the nvidia driver and container runtime already installed that you rent by the hour.

brucethemoose2 3y ago ↗

More task specific compute pools (like Stable Horde) make lots of sense where applicable.