Anonymous
10/13/2025, 7:50:20 PM
No.106876339
>>106876381
>>106877856
Nvidia sisters... Intel just patented AI cubes....
Anonymous
10/13/2025, 8:18:37 PM
No.106876588
Is going to need holes going through it so you can watercool it.
Even during training the intermediate output matrices of a transformer stage are tiny compared to weights. With small batch inferrence, this is even worse. You just don't need gigabandwidth between layers, so it's easier to leave things 2D.
Unless you manage to manufacture the IC like flash, with hundred+ layers on a single die. Then it starts making sense. That's a little harder with logic than flash storage cells.
Anonymous
10/13/2025, 10:30:05 PM
No.106877856
>>106876339 (OP)
That's cool and all but there's no way they'd be able to manufacture that
