Imagine transforming a group of ordinary Macs into a powerhouse AI supercomputer! With the upcoming macOS Tahoe 26.2, Apple is about to make this a reality. But here's the twist: you might not need that shiny new Mac Pro after all.
The Power of Mac Clusters:
Apple's new low-latency feature lets you link multiple Macs via Thunderbolt 5, creating a unified computing beast. This setup is a dream come true for developers and researchers, enabling them to run extensive local AI models with ease. Picture this: four Mac Studios, each capable of handling 512GB of unified memory, effortlessly running the 1 trillion parameter Kimi-K2-Thinking model. And the best part? It's far more efficient than power-guzzling GPU-equipped PCs.
But wait, Mac clusters aren't new. We've seen them before, but they were held back by slower Thunderbolt speeds, especially when hubs were involved, which could throttle speeds to a mere 10 Gb/s. Controversial Opinion Alert: Apple's previous Thunderbolt implementations may have been a bottleneck for high-performance computing enthusiasts. But now, with Thunderbolt 5, Apple promises up to 80Gb/s, a game-changer for data-intensive tasks.
And it's not just the Mac Studio that gets to join the party. The clustering capability extends to the M4 Pro Mac mini and M4 Pro/Max MacBook Pro, too. Developers can build these clusters with just standard Thunderbolt 5 cables and compatible Macs, no special hardware required. Talk about accessibility!
I witnessed the magic firsthand: a cluster of four Mac Studios effortlessly loaded and ran the massive Kimi-K2-Thinking model on an early version of EXO 1.0 by ExoLabs. The most impressive part? This cluster consumed less than 500 watts of power, a fraction of what a typical GPU cluster demands (NVIDIA's RTX 5090, for instance, is rated for 575W and can surge higher).
macOS Tahoe 26.2 also brings exciting news for Apple's open-source MLX project, granting full access to the neural accelerators on the M5 chip, promising a significant boost to AI inferencing. But here's where it gets ironic: the only M5 Mac available, the 14-inch MacBook Pro, supports Thunderbolt 4, meaning it can't tap into the new Mac clustering feature.
Apple Silicon's unified memory and power-efficient design already made Macs a compelling choice for AI tasks. Still, the ability to cluster systems over Thunderbolt 5 takes it to the next level. Sure, a maxed-out Mac Studio with 512GB of RAM is an investment at $9,499 (with the M3 Ultra chip), but it's not the only option. Labs and companies with existing Mac Studios, Mac minis, and MacBook Pros can now consider clustering their current hardware.
The Question: Is this the future of AI computing, or just a niche solution? What do you think about the potential of Mac clusters for AI supercomputing? Share your thoughts below!
