Elon Musk Archive
James Wang
James Wang
@draecomino · 25 ago 2020
Why fp32 for dojo? Will you be leveraging the same chip for in-car inference?
Elon Musk
Elon Musk
@elonmusk
Very different computers are needed for training vs inference. FSD is mostly Int8, but Dojo is de facto FP32 (feels like FP32, but actually FP16 with boring bits truncated) so we don’t need to rewrite all the training software. It should work almost immediately.
04:30 · 25 de agosto de 2020 · 0 visualizaciones
43
71
955
Open on
@jwangARK @wintonARK Very different computers are needed for training vs inference. FSD is mostly Int8, but Dojo is de facto FP32 (feels like FP32, but actually FP16 with boring bits truncated) so we don’t need to rewrite all the training software. It should work almost immediately. — Elon Musk Archive