2 months ago
Noteworthy: this is powered by a new ultra-low-latency userspace TB5 driver offering an ibverbs/RDMA interface over thunderbolt 5[1]. Non-inference HPC workloads could see a massive benefit as well.
[1]: See MLX integration PR at https://github.com/ml-explore/mlx/pull/2808
2 months ago
> Notably, the cluster used less than 500 watts of power
Still startlingly close to microwave oven wattage.
2 months ago
The other notable thing is still waiting for M5 Max and M5 Ultra which should be a big performance increase and maybe more RAM per device, should be amazing for small clusters.