SMT / Hyperthreding


I noticed that some administrators disable Hyperthreding / SMT technology in their HPC environments. I wonder why?

I search around a little bit and I found that the benefit from this technology is strongly dependent on the application itself, but I’m not sure why. May someone clarify what are the pros/cons of enabling/disabling it in scientific HPC cluster? What are the recommendations for DGX A100 systems?

I found this thread in the SLURM users mailing lists that gives a little clarification. But it would be great if someone from NVIDIA could give some comment: