Video wird geladen...
Video konnte nicht geladen werden
Why can't we have arbitrarily large GPU scale-ups? Reiner Pope explains it's partially a matter of wire density:
19,447 Aufrufe • vor 5 Tagen •via X (Twitter)
1 Kommentare

rohan anilvor 5 Tagen
@reinerpope Weakly connected compute island scale up does work (see the hitchhikers blog), taking it to extremes we could touch 1e28 flops, which is pretty neat. Of course coordination of such an enterprise is more complex than training algorithms.

