09-02-2019, 11:30 PM
This screams "workload + tuning value missmatch" to me. And the fact that there are many different types of cards at play here makes this far more difficult to fix. You have 900 series cards, 10 series cards, and 20 series cards all mixed in the same system. This, from my experience, forces you to use non-optimal drivers to maintain cross generation compatibility and can lead to weird tuning problems. Also, the use of x1 risers here will absolutely destroy your card <-> host bandwidth, so attacks that rely on the host for candidates, such as stdin/pipe, dict, etc., will be bottlenecked and thus incapable of proper utilization in a fair number of cases, even with rules used as amplifiers. There are a lot of problems here that need to be fixed or cleared for testing before i would settle on this being a bug in hashcat. Even the system RAM is slightly underspec'ed here, with 33gb of aggregate VRAM and 32gb of host RAM, which does not align with the VRAM<=Host RAM rule that we tend to suggest.