HP Cluster Test Software manual Sgemm Single Precision General Matrix Multiply Test, Memory Test

•A Width of x16 is expected for Gen2 GPUs.

•The Bus ID can be used to identify the physical location of each GPU.

SGEMM: Single Precision General Matrix Multiply Test

•The Trans-Asetting determines whether the matrix A is to be transposed. The default is N.

•ArraySize sets the size of the array to be used. The default is Auto, which means the test will automatically compute the array size. Test results are very sensitive to array size.

•GPU sets which GPU to test. The default is all.

Expected results for Nvidia GPUs: All nodes should report 520 – 550 GFlop/s.

Expected results for AMD GPUs: All nodes should report about 430-440 Gflop/s.

DGEMM: Double Precision General Matrix Multiply Test

•The Trans-Asetting determines whether the matrix A is to be transposed. The default is N.

•ArraySize sets the size of the array to be used. The default is Auto, which means the test will automatically compute the array size. Test results are very sensitive to array size.

•GPU sets which GPU to test. The default is all.

Expected results for Nvidia GPUs: All nodes should report 200 – 250 GFlop/s.

Expected results for AMD GPUs: All nodes should report about 200 Gflop/s.

BandWidth: GPU Bandwidth Test

•Direction sets the direction of the transfers. Available options are htod (host-to-device) and dtoh (device-to-host). The default is htod.

•TransferSize is the number of bytes in a transfer block. The default is 32 GB.

•Iterations is the number of times to repeat the test. The default is 10.

•GPU sets which GPU to test. The default is all.

Expected results for Nvidia GPUs: All GPUs should report 5650-5750 MBs. Values of half the expected range might indicate the GPU is running at Gen1 speed instead of Gen2 speed. This might be caused by a BIOS setting or might indicate a GPU hardware issue.

Expected results for AMD GPUs: All GPUs should report about 3000-3300 MB/s.

Memory Test

NOTE: For Nvidia GPUs only.

This test writes and then reads a pattern to memory and tests for errors. GPU sets which GPU to test. The default is all. All GPUs tested should report zero errors.

Thermal Test

NOTE: For Nvidia GPUs only.

This test reports GPU temperatures for five minutes while a benchmark runs in the background. The GPU temperature should remain below 81° C. GPU temperatures are obtained using the IPMI ipmitool command. It is necessary for IPMI to be installed and enabled for this test to run.

NOTE: The Thermal Test does not report meaningful results for accelerators installed in Workstations (WS490).

24 The accelerator test GUI