"The runs were performed with varying numbers of CPUs. In most cases, 2 CPUs per GPU gave the best performance for a given number of GPUs, therefore all results presented below were done with 2 CPUs per GPU."

Has anyone tried 2 CPUs per GPU???

https://hpc.nih.gov/apps/acemd/benchmarks.html

https://software.acellera.com/docs/latest/acemd3/performance.html

First, these tests were in a Biowolf cluster setup for parallel computing. Not representative of what we run.

Second, are we even sure the acemd3 application in these tests is the same as what we use here at GPUGrid.net?

The application has to be written to use more than one cpu core. I have never seen more than 97-99% of a cpu used for any gpugrid task in BoincTasks.

OTOH, the GW O2MDF tasks at Einstein do use more than 100% of cpu core. I often see up to 110-114% of a cpu core used to compute the gpu task.

In that specific situation, I could see use of a <cpu_usage>2.0</cpu_usage> or a <avg_ncpus>2.0</avg_ncpus> statement in an app_info or app_config for Einstein.