lightweight_mmm
lightweight_mmm copied to clipboard
google
GPU Performance Issues
Hello
We are running lightweightMMM in a Databricks environment using a GPU runtime and we’ve noticed that GPU performance is much slower (10x) than running using CPUs. We are trying to track down the source of the performance degradation and are wondering if others can share their experiences with LMMM on CPU vs. GPU (vs. TPU).
Some background: We are using 3 - 4 years of weekly data (156 / 208 rows) and between 8 and 10 channels and 2-4 extra variables. On CPU we can run the BI-MCMC process with 20,000 warm-up, 20,000 samples on 8 chains in about 20-30 minutes. On GPUs we it takes about 30 minutes to run 1000/1000 on 2 chains.
I’m not sure if the performance we are experiencing is expected or caused by a hardware/software configuration issue.
What are others experiencing? Any help would be much appreciated.
I have same issue when using Lightweight-MMM on Google Colab.
I'm trying to compare Lightweight-MMM and Robyn, using same data: dt_simulated_weekly by Robyn library. dt_simulated_weekly has 208 records.
You might want to check this other issue: https://github.com/google/lightweight_mmm/issues/33
