Hello, i am using two computers for training, one have linux and second windows. These PCs have same HW. When i try to run same training on windows witch batch 4 iam getting CUDA out of memory. On Linux i can have batch 16 and it works fine. To me it seems windows is using too much of GPU memory, is there any way to fix this ? Decrease GPU memory usage by windows ?
Are the GPUs completely empty in both setups or are you using e.g. a display or other applications in the Windows machine?
If they are empty (check it via nvidia-smi), are you seeing the larger memory usage in all models on the Windows system?
1 Like