GPU Benchmark

GPU Benchmark

Benchmarking helps you to evaluate the performance of the tool with standard measurements or similar measurements of its peers.

Running Benchmark in Matlab

Request the GPU node with all GPU cards to avoid conflict at /tmp/GPUBenchReport. You can use: 

(i) Ondemand Desktop (https://ondemand.case.edu/) under InteractiveApps. Make sure that you are using 2 under the "Number of GPU" field and 8gb of memory in Memory field.

OR

(ii) From x2GO or other graphical terminal (refer to HPC visual Access)

srun --x11 -p gpu -C gpuk40 --gres=gpu:2 --mem=8gb --pty /bin/bash          # For Tesla K40m

srun --x11 -p gpu -C gpup100 --gres=gpu:2 --mem=8gb --pty /bin/bash         # For Tesla P100

srun --x11 -p class --gres=gpu:2 --mem=8gb --pty /bin/bash                  # For Markov

Copy the Matlab benchmark utility directory "GPUBench-v2p1" from /usr/local/doc/MATLAB and cd to that directory

cp -r /usr/local/doc/MATLAB/GPUBench-v2p1 .

cd GPUBench-v2p1

Load the matlab module

module load matlab

Run Matlab

matlab &

At the Matlab prompt, run the comamnd:

>> gpuBench

You will see the status bar and once the program is finished, you will see the snapshot similar to the one showed below along with a detailed HTML report on both GPU's performance. 

Fig.  Estimate of the peak performance of GPUs (HPC GPUs and PC in bold) in floating-point operations per second (FLOP/s)

References:

[1] GPU Benchmark