Performance Test

Runtime and Memory Usage

CPU Mode: The GraphST algorithm supports parallel computation, but the number of threads cannot be controlled via parameters. This performance test was conducted on a test node with 112 CPUs and 1TB of memory. Both runtime and memory usage are positively correlated with the number of cells.

GPU Mode: The GPU used is an NVIDIA T4 16G. If the number of cells does not exceed 49,000, the test can be completed normally, and the runtime is positively correlated with the number of cells. For Cellbin/Bin20, the test cannot be completed in either CPU or GPU mode due to excessive memory requirements, resulting in program failure.

Note: In GPU mode with 16GB VRAM, when the number of cells is greater than 46,000 but less than 49,000, you can set export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512 to complete the run. For fewer than 46,000 cells, no additional settings are required.