Hi,
I am confused about the ResNet results from the neural part. I did the same test on my local M1 max machine using the script benchmark_coreml_infer.py The results curve is quite flat.
batch | sam/sec
25 | 487.92
50 | 486.80
100 | 477.55
150 | 470.45
Also, could I know where the GPU part curve comes from since cuda machines are not supporting CoreML.
Many thanks~
Nice work :-)
Here is the results for the numpy benchmark on a 12900k with DDR5 memory using MKL
Results
=======
| datagen | 0.327 |
| special | 0.077 |
| stats | 0.871 |
| matmul | 0.283 |
| vecmul | 0.007 |
| svd | 0.278 |
| cholesky | 0.085 |
| eigendecomp | 3.107 |
Nice work Timothy Liu! I hope you'll continue writing :)
Dave Harris
Thanks Timothy, a really comprehensive breakdown.
Any plans to run additional testing now PyTorch is native? Also, I heard a rumour that M1 GPU FP64 compute was quite strong. Would you know of any benchmarks to corroborate this?