Modern CPUs put a strong emphasis on their AI capabilities and come equipped with an Neural Processing Unit preciscly for that purpose. Would it make sense to use them by this project?

Great question! NPUs are interesting hardware and we've looked into this.

Unfortunately, BOINC doesn't currently support NPU detection — there's no plan_class or resource type for it yet. There is an open issue tracking this: https://github.com/BOINC/boinc/issues/5361 but it's still in the backlog.

On our end, Axiom's architecture already supports multiple compute pipelines (we run separate CPU and GPU experiment tracks today). As soon as BOINC adds NPU detection and scheduling, we'd add a dedicated NPU experiment pipeline alongside the existing ones. NPUs would be well suited for certain experiment types like quantization studies and low-precision inference benchmarks.

So the short answer is: not yet, but we're ready on our side whenever BOINC ships the support.