M5 Max treats SSD as memory
Benchmarks on the M5 Max show SSD being used effectively as extended memory for LLMs, extending Apple’s hardware‑software co‑design playbook for large models at the edge. The tests underline storage‑backed model scaling as a viable lever for on‑device intelligence (x.com).
ANEMLL’s public Results.MD shows the M5 entry in its cross‑generation tables with an estimated 1.3× memory‑bandwidth improvement and roughly 1.2× faster inference times versus the immediate predecessor, with per‑chip benchmark numbers published in the repo. (github.com)) Apple’s product announcement states the new 14‑ and 16‑inch MacBook Pro with M5 Pro/Max starts at 1TB for M5 Pro and 2TB for M5 Max and advertises “up to 2× faster SSD performance” alongside Neural Accelerators in each GPU core. (apple.com)) Independent storage measurements show the latest M5 machines can sustain multi‑GB/s transfers: Fstoppers and other reviewers reported sequential SSD bandwidth claims up to ~14.5 GB/s, while a 4TB review unit recorded ~13.6 GB/s read and ~17.8 GB/s write in disk tests. (fstoppers.com)) MLX/ANEMLL community benchmarks run on 128‑GB M5 Max systems demonstrate that quantized 70–122B models (for example, Qwen3.5‑122B at ~69.6GB in 4‑bit and gpt‑oss‑120B at ~64GB in 8‑bit) can be served on a single laptop and that storage‑backed spill strategies let larger models be probed beyond unified‑RAM limits. (hardware-corner.net)) ANEMLL and associated Hugging Face artifacts show implementation details such as dynamic KV‑cache growth and a “shift‑refill” compaction mechanism that compacts state and spills to storage to sustain very long context outputs (examples cite >24,000 token generation on device). (huggingface.co)) Apple’s ML research and product messaging tie those software techniques to the M5’s Fusion Architecture and per‑core Neural Accelerators, positioning unified memory, much faster NVMe, and MLX as the integrated stack enabling storage‑backed model scaling on MacBook Pro hardware. (machinelearning.apple.com))
