It’s not clear to me. Are these M1 tests running on the CPU, GPU, or neural chip? If the M1 CPU is running at half the speed of a Colab GPU that seems impressive if the software for it is still immature enough that the neural engine/GPU aren’t optimized around.
The M1 Tensorflow-branch doesn't use Neural Engine (ANE). However, it does use the M1-GPU.
The M1 GPU is about 1/3 the speed of a 1080TI card when looking at OpenCL score in Geekbench, but may perform faster than that in some cases, due to the shared memory architecture of the M1.
As I understand it, the ANE has very low precision, which makes it unsuited for training in Tensorflow.
