I built ResNet - like model with Pytorch 1.11. I use a machine with 2 CPU (12 cores), 72 GB memory, no GPU. While trainig the following results were obtained: params forward, s backward, s 76 929 892 2,81 7,1 80 075 818 3,14 7,76 83 322 861 3,32 7,95 86 634 044 3,45 8,47 90 009 367 3,58 8,86 93 516 104 3,47 8,92 96 979 046 3,90 9,23 100 547 209 4,01 9,93 104 179 512 4,27 10,19 107 875 955 4,31 244,46 111 710 028 4,17 255,14 Seems like I achieved limitation of some resource, L3 cache? Any ideas how to resolve it?

Increasing of model size leads to poor backward performance (Pytorch)

I built ResNet - like model with Pytorch 1.11. I use a machine with 2 CPU (12 cores), 72 GB memory, no GPU.

While trainig the following results were obtained:

Seems like I achieved limitation of some resource, L3 cache?

Any ideas how to resolve it?