GPU performance is about 50% slower than benchmarks

Question

Running this benchmark I get 50% slower performance than the author on practically all deep learning sub problems (SINGLE precision and on TRAINING only):

I tested this on a GeForce 1080 GTX Ti and am using PyTorch 1.0.1 with Cuda 10.0.13 (cudnn version 7402) running on an AMD Ryzen 7 1700 Eight-Core Processor (with 16 threads). OS is arch linux. The GPU temperature remains below 80 degrees at full volatile usage.

Would anyone have an idea what causes the lower performance? Can someone else with the same gfx run this and let me know what differences they see?

Answer 1

My guess is that the reduction in performance is due to differences in versions PyTorch. The published benchmark uses torch==0.4.1 (even though the README states something different). You are using PyTorch 1.0.1.

PyTorch 1.0 speed is sometimes slower than lower than 0.4. PyTorch 1.0 can be speed-up by adding torch.backends.cudnn.benchmark = True and compiling from source.

Often times published benchmark are heavily tuned from the "off-the-shelf" version.