I am running the nvbandwidth test on a 8*A100 nvswitch server. My question is, when GP0 reads from GPU1 using cudaMemcpy() and P2P is enabled, can cudaMemcpy() use both the link of pci and nvlink? that is split data into two parts and transfer data via pcie and nvlink separately?
Related topics
| Topic | Replies | Views | Activity | |
|---|---|---|---|---|
| How does “cudaMemcpyPeer” implement? | 3 | 1566 | February 6, 2024 | |
| NVLINK | 3 | 2152 | May 5, 2018 | |
| P2P Transfers Across Single PCIe Switch Fail | 5 | 1538 | April 15, 2024 | |
| DGX-1 using PCIe only instead of NVLink | 2 | 3067 | August 5, 2021 | |
| P2p Bandwidth 150% higher than maximum achievable | 10 | 3197 | April 11, 2023 | |
| Can Unified Memory Migration use NVLink? | 2 | 832 | October 12, 2021 | |
| P2P: How do I know if cudaMemcpy falls back to non-P2P? | 8 | 2659 | October 12, 2021 | |
| The cudaMemcpy speed between two A100Gpus is slow | 1 | 694 | December 22, 2021 | |
| P2P Bandwidth measurements | 1 | 325 | August 6, 2024 | |
| Erratic multi-gpu bandwidth | 8 | 2813 | June 25, 2015 |