Hi, I’m trying to use Taichi in Colab, found it runs very slow (Tesla K80 was allocated for me) even compared to my laptop GPU. I checked the GPU usage on colab during my run, it’s always around 20%. But when I run my code on my laptop GPU, the usage is almost 100%. Does anyone have experience using Taichi efficiently on colab? Thanks very much!
Hi! That’s probably the Colab server cluster throttling your GPU usage. Note that cloud resources are usually shared among several users, i.e. multi-tenancy.
One way to figure out whether this is throttling is to run an arbitrary GPU program, and see if the usage still stays at 20%?
Like @k-ye said, the sharing scheduler might have some problems. The device is way too old, though.
Tesla K80’s core design was first released in the year 2012. It’s actually two K20 packed on a single GPU board. One single chip has ~3TFLops in single precision, so it’s basically a GTX1060.
To fully utilize K80, it needs your Taichi program to 1) utilize 2 GPU chips concurrently and 2) optimize for the legacy Kepler architecture. However, even Nvidia is deprecating the Kepler architecture. So the optimization is not on our roadmap.
As for Colab, is performance your major concern? If you have a GPU at hand, why not run on your own laptop?
Thanks a lot for your answers@haidonglan, @k-ye. I want to test on colab platform first to decide whether paying the pro account Yes,I’m running some fluid simulation program, so performance is a key concern for my application. My laptop can do code debug work, but seems too weak for big cases.
Thank you for your help!