I just checked with the latest version of Spack package that we have with 2 build systems (haven't merged in yet). Seems like the performances match with the ReFrame on interactive run for std-data on cascadelake and cuda on volta. I just double checked my run script and found out that there is an issue with my average of 10 runs script which varied the ncpus after the initial run. That's why I was seeing that much amount of performance gap in some cases. Problem wasn't a a problem of ReFrame or Spack
@kaanolgu