Performance regression on 50-series GPU?

[RKS-labs]

Hi developers,

Thank you for the nice code. As I'm compiling and testing it on 50-series GPU with CUDA 13 (CUDA_ARCH and CFLAGS are modified for sm_120 and c++17), I found the ns/day to be smaller than that of 40-series and CUDA 12. Is the code not designed to run on 50-series, or further optimizations are planned?

[brucefan1983]

Could you provide more information, such as the model size, MD details, and GPUMD version(s)?

[RKS-labs]

Sure. The model is a ~2000 atom cluster. MD is just equilibrating with Langevin thermostat. On 40-series GPUs it's around 4.49584e+06 atom*step/second, and on 50-series GPUs it's around 3.06317e+06 atom*step/second. GPUMD version is 4.7.

[brucefan1983]

Did you use perodic boundary conditions?

Using version 5.0, you willl see both (40 and 50 series) are faster.

The reason that 40 series is faster is because the system is too small. You may need to study a system with > 10000 atoms to see the benifits of 50 series.

[RKS-labs]

I see. PBC is used, and switching off PBC degrades performance. I just downloaded and compiled version 5.0 for both GPUs, and the gap almost closes at this size, the peformance on 40-series and 50-series being around 6e6 and 5.2e6 atom*step/second, resp. Thank you very much!

[brucefan1983]

yes, using PBC will invoke the O(N) algorithm for neighbor list building, otherwise the O(N^2) algorithm is used. So execpt for real 2D materials (favors "T T F"), other systems such as nanoparticles favors PBC for speed.