Silicon Logic

First Report

Article 1 compares llama.cpp against mlx_lm on Apple Silicon, on throughput, perplexity, thermals, and wall and SoC power, using an M5 Max with 36 GB of unified memory.

The report will publish with a pre-registration document and signed run hashes, so the benchmark record is auditable instead of anecdotal.

Methodology Preview

Silicon Logic treats methodology as part of the artifact: run plans are pre-registered, raw outputs are preserved, and claims are tied back to signed measurement records. The methodology paper is forthcoming; the v0.9 pre-print stub reserves the public URL and names the benchmark harness, measurement boundaries, and reporting rules before the first review lands.