Interesting, both look like certain quantisation of the weights. I bet, related reduction in memory bandwidth would be even larger than win from the simplified computation though. That said, why do you think Tensor Cores won't last long? :)
-
-
-
Well mostly because they are basically there to do "fused multiply add" and that's it. That operation is the most effective tactic available now but likely not as much in the far (maybe near?) future as those papers for example hint.
- Još 2 druga odgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.