The GPU does not improve the performance as in the example shown in Matlab

Hello,

As part of a task we are currently performing, we are required to use a GPU to speed up the calculations.

The tests are performed by using NVIDIA T1000 GPU (4 GB).

In order to test the runtimes achieved using the GPU, we ran this code in MathWorks:

Improve Performance Using a GPU and Vectorized Calculations - MATLAB & Simulink (mathworks.com)

There is a gap between the results we got and the results in the example (the results are attached).

In the unvectorized case (blue) the run times we obtained by using a GPU are significantly higher than the run times obtained by using a CPU, although in the example similar times were obtained in both modes.

In addition, in the vectorized case (orange) the difference we obtained between the CPU and the GPU is significantly lower compared to the difference obtained in the example.

We can conclude that the effect of the GPU in our case is not so significantly felt, and this is also evident in other more complex examples that we tested.

I would like to emphasize that the latest versions of the GPU and CUDA have been installed, and it seems that there is no memory problem that prevents these operations from being performed in the optimal way.

What could be the reason for this?

Thank you,

Nadav

[Damian Pietrus]

Out of curiosity, do you get the same types of results when re-running the code a few times in a row?

[Avi]

Yes, I get exactly the same results.

[Abhishek Kumar Singh]

@Avi This is a quite unique instance. I ran it on RTX 4080 and got the expected results.

Can you please share output of gpuDevice when executed in command window?

Answer 1

[Ravi]

Hi Avi,

The graphics processing capabilities of the NVIDIA T1000 currently in use do not quite match the more robust performance parameters of the NVIDIA RTX A5000 as highlighted in the provided example. For example, the T1000 has a peak single precision performance of 2500 GFLOPS, while the RTX A5000 has an estimated peak performance of 21,700 GFLOPS. The disparity in performance is quite pronounced and may account for the observed variance in acceleration between the vectorized version when conducting a comparative analysis of CPU versus GPU performance across the two graphics processing units.

Hope this answers your question.

Thanks,

Ravi