The Higgs training runs for LightGBM take the same amount of time for me on both GPU and CPU - 26 seconds. Logs confirm that GPU run is using GPU (transferring data to GPU etc.)
https://lightgbm.readthedocs.io/en/latest/GPU-Tutorial.html
Went through tutorial to install LightGBM for GPU. Installed fine and was able to run GPU training to confirm.
GPU and CPU specs below for comparison.
Processor (Skylake X; Latest Generation)
8-Core 3.60 GHz Intel Core i7-7820X
VS
NVIDIA RTX 2080 8 GB
Note: accepted all configuration defaults as per tutorial as this is the benchmark. Can play around with them but that might be defeating the point.
Anyone tried similar on RTX?
Performance benchmarks https://lightgbm.readthedocs.io/en/latest/GPU-Performance.html are using GTX but indicate that any recent NVIDIA card should indeed work. The benchmark page warns against Kepler cards. But RTX 2080 is Turing and seems to support hardware atomic operations. https://en.wikipedia.org/wiki/CUDA My limited knowledge suggests it should all be good.