Why opt for an upgraded Mac Pro when linking various Mac Studios can form a single powerful setup? Apple's forthcoming macOS Tahoe 26.2 update includes an innovative low-delay tool that enables connecting multiple Macs through Thunderbolt 5. This offers developers and scientists an effective method to assemble robust AI systems capable of handling enormous on-device models. As a result, a group of four Mac Studios, each supporting up to 512GB of shared memory, can process the 1 trillion parameter Kimi-K2-Thinking model with greater efficiency compared to PC setups relying on energy-intensive GPUs.
Previous attempts at Thunderbolt-based Mac clusters faced constraints from reduced data rates, particularly when hubs were involved, dropping performance to 10 Gb/s. The latest Apple innovation supports the complete 80 Gb/s capacity of Thunderbolt 5. This clustering function extends beyond the expensive Mac Studio to include the M4 Pro Mac mini and M4 Pro or Max versions of the MacBook Pro. Building such setups requires no extra equipment from developers, only Thunderbolt 5 cables and suitable Macs.
During a demonstration, a setup of four Mac Studios successfully loaded and executed the large Kimi-K2-Thinking model using an initial build of ExoLabs's EXO 1.0 software. Impressively, the entire cluster consumed under 500 watts, making it about ten times more power-efficient than standard GPU configurations. For reference, NVIDIA's RTX 5090 GPU is specified at 575W, though actual usage can exceed that.
Additionally, macOS Tahoe 26.2 will unlock full utilization of the M5 chip's neural engines for Apple's open-source MLX framework, significantly accelerating AI inference tasks. However, the current sole M5-equipped device, the 14-inch MacBook Pro, is limited to Thunderbolt 4 ports, preventing it from participating in this enhanced clustering feature.
The architecture of Apple Silicon, with its integrated memory and efficient power consumption, has long positioned Macs as a strong option for intensive AI tasks. The new option to interconnect systems via Thunderbolt 5 could prove especially appealing for those dealing with expansive models. Bear in mind, a Mac Studio configured with 512GB of memory begins at $9,499 featuring the M3 Ultra processor, representing a premium choice. Nonetheless, research facilities and businesses owning existing Mac Studios, Mac minis, or MacBook Pros might leverage their current inventory for clustering.