I am training my tensorflow model on google cloud run, using TensorFlow CPU. However, I am only getting around 65% CPU usage (Using 4 CPUS) and the model is taking a little while to run. I was wondering if anyone knows any optimisation that is possible for TensorFlow CPU , especially on cloud run. I couldn’t find any articles about this
Thank you
hello, welcome, why the CPU? by intellectual challenge? no GPU? a GTX1650 doesn’t cost much… TF is GPU optimized, certainly Keras is distributed in CPU mode by default on Debian, but that’s not a reason
If you are on an Intel/xeon instance you can install Intel optimized packages:
Thanks I’ll take a look at this
Only difference and the latest TF OneDNN is built in and ready to go so its down to you.
GTX 1650 really is a bum steer as 4GB for some models doesn’t make the grade and still £200 which is just crazy for what it is.
Its time though and my RTX 3050 just completed a DTLN model in an hour that would of taken 10 hours on a I7 6700 and that is just how things are as even with Intel® AVX2 unless you go for the latest and greatest with AVX512 still in comparison to a decent GPU they are relatively slow but without the stupid $
AVX512 is only a bit faster than mine and even though a low end RTX or old 1070 your not going to get anywhere near the 10x speed they can provide.
1 Like
like what the 3050, is not such a GPU, have you thought about making yourself a cluster based on RTX 3050? there are plenty of open source solutions…