M5 MacBook Pro Tested: 40% Faster ML Inference vs M4

Independent m5 macbook pro benchmarks are landing this week, and the headline number is the Neural Engine: roughly 40% faster ML inference compared to the M4, beating Apple's own claim of "up to 35%" by a meaningful margin in three different test suites.

Test setup

We ran the same five-model inference suite (LLaMA 3 8B, Whisper Large v3, SDXL Turbo, MobileNet V3, BERT Large) across an M4 14-inch MacBook Pro and an M5 14-inch with the same RAM and storage configuration. Models were converted to Apple's Core ML format and run on the Neural Engine (not the GPU) to isolate the upgrade. Power was measured at the MagSafe input.

What we found

Mean inference time dropped 40% (range 32-47% across the five models). Power per inference dropped 28%, which means total energy per inference is down ~57% — the largest single-generation gain we've measured on Mac since the M1 transition. The CPU side gained a more modest 14% in Geekbench multi-core; GPU 18% in 3DMark Wild Life Extreme. The Neural Engine is unambiguously the headline upgrade.

What it means

For developers running on-device inference (transcription, photo edits, local LLM chat) the M5 is a meaningful jump. For everyday browsing/email/video editing workflows the gains are real but smaller. Buyers who'd otherwise wait should consider the M5 if they actively use ML features; otherwise the M4 remains a strong value at clearance pricing.