M3 Ultra achieves 45.8% performance gain running 27B model with CPU+GPU+ANE offload
AIFlow_ML · x · 2026-08-20
A developer benchmarked a 27B model on an M3 Ultra, achieving a net 45.8% performance increase (517.9 PP tok/s) compared to GPU alone by utilizing 3 CPU cores alongside the GPU and ANE. A new tuner was also introduced, which balances 5 dimensions in minutes instead of over 10 minutes for the previous 2-dimension version.
More from Infra
- 4th generation cryostat unveiled with 250+ photonic chips — jwt0625 · 2026-08-20
- Ethernet copper costs as low as $0.10-0.25/Gbps — jwt0625 · 2026-08-20
- Nvidia's Sustained Performance Limited by Power Delivery: 9.7 PFLOPS — nabla_theta · 2026-08-20
- ML Researcher: Compute Shortage, Not Ideas, Slowing Down Progress — A_K_Nain · 2026-08-20
- Llama.cpp restarts kill long context prefills: proposal for auto KV state persistence — Dazzling_Equipment_9 · 2026-08-20
- YC-backed Atomarine builds floating data centers to solve the compute crunch — ycombinator · 2026-08-20