3 ms·
One thing about AI research now is a lot of stuff seems to be built on pretty big local GPU clusters or pretty expensive AWS stuff. I had an idea the other day
by ilaksh 10y ago
One thing about AI research now is a lot of stuff seems to be built on pretty big local GPU clusters or pretty expensive AWS stuff.
I had an idea the other day that could make that more convenient/affordable. The trick of course is coming up/applying different/new algorithms that don't rely on fast synchronization. Ideally that would be some kind of AI language that could describe/compress complex models/filters/parallel functions/etc. (https://github.com/runvnc/ailang https://github.com/runvnc/ailang) Then the nodes could exchange 'knowledge' in that special language.
Combine something like turbo.js with a webrtc rpc swarm. Use rpc to pull out of task queue. So we can do GPGPU processing on web-based clusters in the browser.
Can just use public gists with 'gpgpucluster' at end of description. Then use requirebin. Also use webtorrent/instant.io
for imagenet example. Need to break imagenet up into a ton of little torrents. Then make an example that maybe just
searches for certain classes like 'cat'. Example doesn't need to do actual machine learning, just store images in graphics memory
pull them out when they search. And send them back. Maybe just use a subset of imagenet so that multiple clients
can share the dataset and so work together to answer queries.
EDIT: Thanks for the downvotes! I thought it was an interesting idea. But then again, the more interesting I think an idea is, the more likely it is to be downvoted all to heck. Hard for me to suppress my desire to share my ideas even though they are not appreciated.
- Cacti 10y agoIt's hard to tell what you're asking for. There is no getting around the amount of computation required, and there are already plenty of systems to distribute workload the old fashioned way. In addition, while many projects rely on clusters of GPUs, a great deal of cutting edge research can be done on one or two decent commercial grade GPUs. A single new NVIDIA 1070 (and even better, the coming 1080 Ti) is cheap ($400) and is comparable to many good supercomputers in 2000, for example. An AWS with a K80 is only about 30 cents an hour at spot prices... after a week of training you could have yourself a state of the art result that wasn't even possible a mere few years ago. Not too shabby! Things have never been cheaper.