CUDA GPU Plugin

[H in OH]

Is there a plugin for BIT to exercise CUDA based GPGPUs? If performing a thermal load test on a system with Nvidia Tesla GPGPUs, the GPUs really need to be brought up to full power along with the CPUs for a fair test.

[David (PassMark)]

There is a "GPGPU" test built into V8 of the software. So there is no need for a plug-in.
(GPGPU = General Purpose Graphics Processing Unit)

This test works using DirectCompute. Which is the same concept as CUDA, except DirectCompute is more general, in that it works on Intel, nVidia and AMD GPUs.

From the BurnInTest help file,

The GPGPU test exercises and verifies the GPU's General Purpose computing operations (rather than 3D type operations). TheGPGPU tests are a bit like the CPU tests, but the tests are run on the GPU. The GPGPU test load is user selectable.


The GPGPU test runs on all detected GPUs (to a maximum of 8 GPUs). The GPUs that are being tested are listed in theGPGPU test window. The numbers displayed in the test window for this test represent how many millions of GPGPU operations have been performed and verified.


Each different GPGPU test is run for a defined period of time. Integer maths and floating point maths tests are run together. After all tests have been run the cycle count is incremented. The duty cycle (and the hardware capability) determines how many operations can be processed during the test cycle.

[H in OH]

There is a "GPGPU" test built into V8 of the software. So there is no need for a plug-in.
(GPGPU = General Purpose Graphics Processing Unit)

This test works using DirectCompute. Which is the same concept as CUDA, except DirectCompute is more general, in that it works on Intel, nVidia and AMD GPUs.

From the BurnInTest help file,

The GPGPU test exercises and verifies the GPU's General Purpose computing operations (rather than 3D type operations). TheGPGPU tests are a bit like the CPU tests, but the tests are run on the GPU. The GPGPU test load is user selectable.


The GPGPU test runs on all detected GPUs (to a maximum of 8 GPUs). The GPUs that are being tested are listed in theGPGPU test window. The numbers displayed in the test window for this test represent how many millions of GPGPU operations have been performed and verified.


Each different GPGPU test is run for a defined period of time. Integer maths and floating point maths tests are run together. After all tests have been run the cycle count is incremented. The duty cycle (and the hardware capability) determines how many operations can be processed during the test cycle.

BIT 8.0 was able to detect standard GPUs like the Nvidia QUadro series. It did not detect a pair of Nvidia Tesla K20 GPGPUs that are in TCC mode for computation only and no video output.

Thanks.
---H

[Spiff]

I ran into a similar issue when trying to test a Nvidia Tesla K80 card, I had to switch the card from TCC to WDDM mode using the nvidia-smi tool.

Is there something that can be done to fix the GPU detection in the GPGPU test so it can see Tesla cards that are in compute only mode?

When running just the GPGPU test on the K80 the actual load on the card was as follows.

80% duty cycle ~7% load
95% duty cycle ~75% load
100% duty cycle ~93% load

When running the CPU tests and the GPGPU tests at the same time the GPU load would only be at about 7% with the CPU @ 80% duty cycle and the GPGPU @ 100% duty cycle.