Bit-exact benchmark suite for neuromorphic inference
Anulum Institute / SC-NeuroCore
Challenge
Neuromorphic and stochastic-computing kernels are often compared by headline speed, but speed without bit-exact parity evidence is meaningless: a faster kernel that silently changes numeric behaviour invalidates the science it supports. The field lacked a multi-language, multi-backend benchmark suite where every claim is rendered at its true boundary.
Approach
SC-NeuroCore built a benchmark databank covering eight inference capabilities across more than seventy measured cells, spanning Rust, Mojo, Julia, Go, Numba, CuPy, JAX, and Torch backends. Each cell records wall-clock latency, numerical parity against a canonical reference, and the exactness class required by the platform contract. Results are aggregated by scpn_studio_platform.benchmark.aggregate, sealed with studio.publication-seal.v1, and appended to the Anulum transparency log.
Results
The databank surfaces honest inversions: Julia is orders of magnitude slower on gpfa and dimensionality than expected; Python outperforms the Rust port on dimensionality for this workload; and the declared dispatch order rust→mojo→julia→go is contradicted by the studio's own measurements in all eight capabilities. These contradictions are pinned as terminal evidence rather than laundered.