M5 Ultra Tested: Memory, Storage and Local LLM Speed vs M3 Ultra

When Apple’s new M5 Ultra Mac Studio landed next to my M3 Ultra, my first question was simple: do the claims of up to 4.3× local AI, up to 2× storage and 1.2 TB/s bandwidth hold outside the slide? The base system starts at $5,499 , while the 80-core GPU, 256 GB unified memory and 8 TB storage config on my desk reaches $14,299, with a 512 GB option due next month. Apple loaned the M5 unit, the M3 is my own purchase, so I could run both under identical Llama.cpp and MLX builds with identical model files .

CPU and Everyday Performance: First Measured Gaps

On Speedometer the M3 Ultra scored 47 versus 60.2 on M5 Ultra; a single M6 core is faster in isolation, but the Studio chassis still shows this generational lift. For developers, a large .NET build with 100k namespaces and classes fell from 71.7s to 52.4s, and a Python Mandelbrot style interpreted workload dropped from 10.25s to 5.8s, roughly 1.9× . The machine is overkill for pure coding, yet the multicore and interpreter gains are clean and repeatable.

Storage beats the marketing line. Sequential copy with AmorphousDiskMark gave roughly 6.5 GB/s read and 2.7 GB/s write on M3 Ultra versus 14.9 GB/s read and near 20 GB/s write on M5 Ultra; random small-file copies were also well beyond 2×. When you need to pull a 140 GB model from disk into memory , that delta directly shortens load time, and both drives were 8 TB so the comparison is apples-to-apples.

Memory Bandwidth and the AI Architecture Shift

On paper Apple lists 819 GB/s for M3 Ultra and 1.2 TB/s for M5 Ultra. With STREAM Triad , a CPU-centric test, I measured 312.9 GB/s vs 583.6 GB/s; the GPU micro-benchmark , which matters for AI, gave 724 GB/s vs 1,039 GB/s , about 1.4× real bandwidth on M5. That ratio tracks closely with the uplift seen in token generation, which is bandwidth-sensitive.

The core architectural change is that M3-era GPU cores were plain graphics cores, while each M5 GPU core now houses a dedicated neural accelerator for matrix multiply . That is why the same Llama.cpp binary reports `has_tensor = false` on M3 and `has_tensor = true` on M5, unlocking a Metal path tuned for the M5 family . MLX , Apple’s own stack, is often fastest overall, but this hardware hook explains M5’s sharp prompt-processing jump in Llama.cpp. All runs used identical builds and model weights on both machines.

The distinction between prompt processing (PP) and token generation (TG) matters here. PP — reading your prompt — leans on compute and GPU core speed , while TG — writing the answer — leans on memory bandwidth . M5 Ultra lifts both, but PP benefits disproportionately from the per-core accelerator, which is why long prompts pull away.

LLM Numbers: How Much Faster Per Model

On the large mixture-of-experts DeepSeek V4 Flash 284B , TG rose from 37 to 53 tok/s, about 1.5× , while PP leapt from 483 to 1,485 tok/s, roughly 3× . On GPT-OSS 120B , TG moved from 90 to 130 tok/s and PP from about 1,300 to 3,200 tok/s; feeding 32k new tokens on top of 64k already in context took 80s vs 56s, again near 1.4× . The pattern is consistent: TG follows bandwidth, PP follows the new accelerators.

On a smaller but dense Qwen3 27B , where every token touches all weights, the load is heavier yet still revealing. TG went from 37 to 54 tok/s, 1.47× , while PP jumped from 430 to 1,800 tok/s, over 4× . A realistic 14k-token code prompt equivalent to a handful of files saw PP fall from 33s to 8s, 4.2× , confirming Apple’s up-to-4× claim right where it matters — long, file-fed contexts. The curve is prompt-length dependent: about 1.6× at ~350 tokens , 3.4× near 1,700, and the 4× threshold around 4,500 tokens , so chat feels modest while agent workloads shine.

Power and thermals tell a more tempered story. During generation the chip telemetry showed 36W on M3 vs 45W on M5 , so tokens per watt rose only ~12%; at idle it was ~11–12W vs 8–10W, slightly lower on M5. Under sustained full GPU load, M3 peaked near 200W while M5 pushed beyond 400W , running visibly hotter around 43–44°C and audibly louder on mic. Efficiency improved, absolute power and heat did not.

One more note from the lab bench: Llama.cpp’s Metal path now branches for the M5 family while MLX’s per-core tuning keeps evolving, so the same silicon has two different peaks and leadership flips with model, context length and quantization.

Capacity today fit comfortably within 256 GB , but the coming 512 GB option will unlock still larger models and faster multi-model setups. With eight concurrent requests , aggregate throughput rose from 75 tok/s to 134 tok/s, showing the gain persists under concurrency. A deeper dive into MLX versus Llama.cpp and 8-bit variants is next; the first-look takeaway is clear — ~1.5× TG across models and 2–4× PP depending on prompt length , with coding agents as the prime beneficiary.