Verify the cause of species-dependent CPU performance scaling

Verify using hardware-counter measurements whether the increase in Symmetrix-XL per-atom inference cost with the number of chemical species is caused by the larger spline-table working set and reduced cache reuse.

Background

Symmetrix-XL exhibits increasing per-atom CPU inference cost as the number of chemical species and associated radial spline tables grows. The paper suggests that a larger spline-table working set and reduced cache reuse may explain this trend, but explicitly states that this interpretation has not been verified with hardware counters. Establishing the underlying performance mechanism would clarify whether the observed scaling is attributable to cache behavior or to other computational or memory-system effects.

References

A larger spline-table working set and reduced cache reuse are plausible contributors, although this interpretation has not been verified using hardware counters.

— Train for Accuracy, Execute at Scale: Architecture-Preserving Inference for Equivariant Atomistic Foundation Models  (2610.01036 - Fu et al., 1 Oct 2026) in Section 5.3, “CPU performance”