[account deleted]

You were right — some of our experiment scripts were triggering multi-threaded BLAS operations (numpy's internal linear algebra routines like eigenvalue decomposition), which would spawn extra threads even though each task was declared as 1 CPU to BOINC.

We identified the root cause and deployed a fix in v6.34 (all platforms, released today). The binary now forces OMP_NUM_THREADS=1, MKL_NUM_THREADS=1, and OPENBLAS_NUM_THREADS=1 at startup, so numpy's BLAS layer is locked to a single thread per task. We've confirmed on a test machine with 80 cores that every running task shows exactly 1 thread.

BOINC's CPU accounting should now be accurate — each task truly uses 1 CPU, so the scheduler can correctly calculate concurrent task slots based on your CPU usage preferences.