4 ms·
Even without any compression, if you use INT8 for the weights, instead of FP32, you will get 4x size reduction.
by p1esk 10y ago
Even without any compression, if you use INT8 for the weights, instead of FP32, you will get 4x size reduction.
- billconan 10y agoyes, that's one trick they did in the paper. they also do pruning, weights close to zero are removed. then they implemented a gpu sparse matrix multiplication kernel.