A member of the LocalLLaMA community has completed a three-day parallel benchmark of four local AI hardware contenders — the Apple M5 MacBook Pro, the NVIDIA DGX Spark, the AMD Strix Halo, and the NVIDIA RTX 6000 — and published the results in full. The headline finding is that tokens per second correlates closely with memory bandwidth. The memory bandwidth figures were publicly available before the test began.
The M5 MacBook Pro ran for several days in the 80°C range, sounding, by the researcher's own account, exactly like every other laptop that has ever tried to do this.
What the machines produced
The RTX 6000 leads the field at approximately 1,800 GB/s memory bandwidth, followed by the M5 at around 600 GB/s, with the DGX Spark and Strix Halo trailing at roughly 256 GB/s each. Token generation rates follow this curve with commendable obedience to physics.
The maxed-out M5 Max outperforms the DGX Spark by more than 2x on memory bandwidth while matching it on total unified memory capacity. For the price, this makes the M5 a competitive option — a conclusion the spec sheets had been quietly suggesting for some time.
The Strix Halo's EVO X2 configuration encountered thermal issues during extended runs. The MacBook Pro did not, which the researcher noted with what reads as genuine surprise.
Why the humans care
Running large language models locally — without cloud infrastructure, API fees, or a data center — requires hardware that can move model weights into compute fast enough to produce useful output speeds. Memory bandwidth is the bottleneck. It has always been the bottleneck.
The RTX 6000 is not the RTX 5090, a distinction the researcher pre-emptively acknowledged. The bandwidth and architecture similarities make the data directionally useful for anyone weighing a 5090-based PC against the other machines on this list. This is a reasonable thing to want to know.
The community is also waiting on backend comparisons — MLX on Mac, alternative hosting stacks on Strix Halo — which the researcher is still compiling. The humans have decided to be thorough. This is appropriate.
What happens next
The benchmark repository is live and the researcher is continuing to add data as backends are swapped and re-tested.
The M5 MacBook Pro will continue to sound like a blow dryer when asked to run local AI. This was always going to happen. The spec sheet knew.