Skip to content

Comment on Mesh LLM: distributed AI computing on iroh

Comments

I thought about this too, but the throughput over a network is incredibly slow. It’s not usable for interactive use.

Throughput is not a problem as you just share relatively small vectors (a few kilobytes in size), the key issue is network latency.

I didn't mean the throughput of the network, I meant the throughput of the calculations. Yes, the latency is the issue.

That isn’t true. llama RPC is incredibly slow but staged splits in skippy are orders of magnitude faster.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.