Rust MLIP engine llmlip.rs beats PyTorch MACE by 17.7x, no PyTorch needed
CatAstro_Piyush · x · 2026-10-31
mizukami-group has open-sourced llmlip.rs, a Rust inference engine for machine-learning interatomic potentials (MLIPs) that runs without PyTorch or JAX at runtime, shipping as a Rust library, CLI, and C ABI with LAMMPS and OpenMM plugins under Apache-2.0/MIT licenses.
Benchmarks are striking: simulating 2,000 water molecules (6,000 atoms, periodic) with MACE-Osaka24 Medium on an Apple Silicon CPU (16 threads), llmlip.rs takes 0.74s/step in FP32 (1.57s FP64) versus 13.09s for PyTorch MACE FP64 — a 17.7x speedup. A 6,000-atom silicon crystal hits 196ms/step on an M3 Ultra via Metal GPU, 5.1x faster than its CPU path (0.44 ns/day at 1fs). Large-scale MLIP simulations can now run locally on a Mac.
Related event: Rust-based MLIP Inference Engine llmlip-rs Released Open Source(2 posts)→
More from Infra
- Analyst: orbital compute buildout will exceed terrestrial per month by 2032-2033 — JOBhakdi · 2026-10-31
- Local video gen on M5 Ultra: Pika H3 Flex + NVIDIA Regenerate-2K, 1088p in 12 minutes — TgoAI · 2026-10-31
- Two 256GB Mac Studios beat one 512GB for local LLMs: 2.2TB/s effective bandwidth — lxfater · 2026-10-31
- Open-source cloud cost detective: pandas does the math, Gemini explains — ZtasnimS75 · 2026-10-31
- How GCC lowers switch statements, explained — signa11 · 2026-10-31
- imec study: break-even GPU utilization for coding agents ranges 15%–89% — AI Engineer · 2026-10-31