A surprising fraction of machine learning performance engineering is figuring out how to profile the system component that shouldn't be your bottleneck but somehow is.

May 20, 2022 · 2:23 AM UTC

2
20
1
202
Replying to @gdb
The problems start even before one gets to this stage. The first step is what is causing the performance bottleneck? The HW, connections (+switches), HW components or even the order of the plugged in components. Next, comes the software stack.
Replying to @gdb
Now it is but future it is not