Optimize GPU Performance
Crusoe Cloud is currently in private beta. If you do not currently have access, please request access to continue.
Across all of the NVIDIA-based Crusoe accelerated instances it is possible to further optimize the performance of the GPUs by leveraging a clock locking mechanism to reduce latency and maximize performance of the workloads. With the NVIDIA driver installed apply the following settings:
Ensure that the GPUs are in persistent mode:
sudo nvidia-smi -pm 1
Graphics Clock Locking - To reduce clock switching latency and ensure that the maximum SM clock is available across all execution kernels in your code. You can set all GPUs to the maximum value:
nvidia-smi -i 0 --query-supported-clocks="gr" --format=csv,noheader | head -n 1 | awk '{print $1}' | xargs sudo nvidia-smi -lgc
Memory Clock Locking - To reduce clock switching latency and ensure that the maximum memory clock is available across all execution kernels and DMAs in yoru code. You can set all GPUs to maximum value:
nvidia-smi -i 0 --query-supported-clocks="mem" --format=csv,noheader | head -n 1 | awk '{print $1}' | xargs sudo nvidia-smi -lmc
You can confirm that the settings were applied based on the clock SMs/Memory values below:
| GPU | Max Graphics Clock (Mhz) | Max Memory Clock (MHz) |
|---|---|---|
| NVIDIA A6000 | 2100 | 8001 |
| NVIDIA A40 | 1740 | 7251 |
| NVIDIA A100 40GB PCIe | 1410 | 1215* |
| NVIDIA A100 80GB PCIe | 1410 | 1512* |
| NVIDIA A100 80GB SXM4 | 1410 | 1593 |
| NVIDIA H100 SXM5 | 1980 | 2619 |
*Setting the memory clock through the locking mechanism is not supported
For workloads that can tolerate memory errors (ie. Graphics targeted workloads) or if your code has an out-of-band error correcting mechanism. You can maximize performance of your code by disabling ECC. To disable ECC run the following:
sudo nvidia-smi -e 0
which will disable ECC for all GPUs in the instance, a reboot of the instance will be required to take effect.
MIG - To partition certain GPU types into multiple instances for running different, isolated workloads, you can run
sudo nvidia-smi -i 0 -mig 1
# -i 0 partitions gpu with index 0
For more information and to see which Nvidia GPUs support this, see this link.