Four Mac Studio Units Clustered with 1.5 TB Memory Run AI That Does Not Fit on a Single GPU
Apple
NVIDIA
Exo
Moonshot AI
DeepSeek
Alibaba/Qwen
A journalist combined four Mac Studio units with M3 Ultra into a cluster with a total memory of 1.5 TB using RDMA technology via Thunderbolt 5. The cluster ran massive AI models such as Kimi K2 Thinking (over 600 GB), which cannot be loaded on a single computer. The project was implemented using the open-source tool Exo 1.0.
Apple provided the author with four Mac Studio units on M3 Ultra for testing RDMA over Thunderbolt — a new feature of macOS 26.2 that allows combining the RAM of multiple computers into a shared pool. The total unified memory of the cluster amounted to 1.5 TB, and the cost of such a set was around $40,000 (the computers were borrowed for tests). The Exo 1.0 tool was used for task distribution. Thanks to RDMA, memory access latency dropped from 300 microseconds to less than 50 microseconds. The cluster consumes under 250 watts and runs almost silently. The two lower Mac Studio units have 512 GB of memory and 32 processor cores each ($11,699 each), while the two upper units have half the memory ($8,099 each). In the HPL benchmark, one Mac Studio delivered 1.3 teraflops, while the cluster of four achieved 3.7 teraflops. Activating RDMA required manually entering a command in recovery mode. The cluster enabled running models weighing over 600 GB, such as Kimi K2 Thinking. Performance in the Qwen3 235B benchmark was 32 tokens per second on the full cluster.
Source: Habr — хаб ML —
original