3 ms·
Hmm, I'd expect that once you've got your weights loaded machine learning would need much less bandwidth/flop than graphics does. Is that incorrect?
by Symmetry 10y ago
Hmm, I'd expect that once you've got your weights loaded machine learning would need much less bandwidth/flop than graphics does. Is that incorrect?
- LogicFailsMe 10y agoApples and oranges. And as it turned out, it's an 8-bit fixed point processor. Useless for training without heroic handholding, limited use for inference. Google's PR machine does it again.