A Compute and Communication Runtime Model for Loihi 2

Jan 15, 2026 by Jonathan Timcheck, Alessandro Pierro, S. Shrestha (arXiv.org)

DOI 10.48550/arXiv.2601.10035

We built the first max-affine, multi-dimensional roofline runtime model for Loihi 2 that quantitatively captures compute and noisy NoC communication costs from microbenchmarks, then showed it predicts real runtimes (r≥0.97) and exposes area-runtime tradeoffs and scaling laws for matrix-vector layers and a QUBO solver. If you care about squeezing latency out of neuromorphic chips, this gives a lightweight, predictive tool to guide spatial mapping and algorithm design.

source S2, openalex