Almost every year a new type of GPU is released. Our research community is highly diverse and runs both single and double precision GPU applications. We would like to compare various kinds of GPUs and see how much a new type of GPU will improve the performance of applications that are run on our cluster. Are there any suggestions on which software applications, that are commonly run in a cluster environment, would be the best for this purpose? We are especially interested in Chemistry/Material Science applications and possibly some Machine Learning applications.
The benchmark suites listed below feature GPU applications that might help with the Chemistry/MatSci application search :
[1] Unified European Applications Benchmark Suite, repository available here, report describing the same available here
[2] UK-HPC benchmark suite, repository available here, reports listed in the repository.
A few ideas if you want to use customized versions of HPL:
If it is an NVIDIA GPU, you can get a version of HPL that is GPU enabled if you are a registered CUDA developer.
Also, https://github.com/davidrohr/hpl-gpu looks like a more platform agnostic version of HPL that supports OpenCL. I’ve not personally tried it though.
I don’t have experience with Chemistry or ML application benchmark testing, but HPL runs between different kinds of GPUs may at least give you some relative comparisons.