3 ms·
The key to doubling AI operations is changing the architecture significantly, which sadly makes it harder to use the platform itself as well. After CUDA we had
by xiphias2 6y ago
The key to doubling AI operations is changing the architecture significantly, which sadly makes it harder to use the platform itself as well.
After CUDA we had to use Tensor cores, later float16, now sparce matrix multiplications that work because deep learning uses lots of RELUs.
- mikorym 6y agoCD players are a more legacy example of hardware doing magic, in this case, a whole lot of XORs. [1] [1] https://youtu.be/fBRMaEAFLE0?t=651 https://youtu.be/fBRMaEAFLE0?t=651 "Free with every CD, a Cray supercomputer"