RE: LeoThread 2026-02-23 19-58

avatar

You are viewing a single comment's thread:

If a model can be burned into a thermoputer and each forward pass run as a single relaxation—taking output tokens and re-clamping the inputs—a thermoputer could cycle nearly a billion times per second

leofinance
0
0
0.000
3 comments
avatar

That implies a theoretical throughput on the order of 100 million tokens per second, with energy use reduced by roughly a billion-fold

0
0
0.000
Reply
avatar

The main bottleneck for AI speed is loading weights into memory
Copper I/O today isn't fast enough, and optics only brings limited gains
Future chips will address this by embedding weights directly into the compute substrate

0
0
0.000
Reply
avatar

Talaas is an early step toward that approach

If those hardware directions succeed, the coming period will be very interesting

0
0
0.000
Reply