You Can Turn A Cluster Of Macs Into An Ai Supercomputer In Macos Tahoe 26.2

Who needs a revamped Mac Pro erstwhile you tin conscionable move respective Mac Studios into a unified computing system? With nan upcoming macOS Tahoe 26.2 release, Apple is introducing a caller low-latency characteristic that lets you link respective Macs together utilizing Thunderbolt 5. For developers and researchers, it's a perchance useful measurement to create powerful AI supercomputers that tin tally monolithic section models. That allows 4 Mac Studios, which tin each tally up to 512GB of unified memory, to tally nan 1 trillion parameter Kimi-K2-Thinking exemplary acold much efficiently than PCs pinch power-hungry GPUs.

While we’ve seen Thunderbolt Mac clusters before, they were constricted by slower Thunderbolt speeds, particularly if they required a hub (which could trim speeds to 10 Gb/s). Apple’s caller characteristic allows for nan afloat Thunderbolt 5 connectivity of up to 80Gb/s. The clustering capacity besides isn't conscionable constricted to nan pricey Mac Studio, it will besides activity pinch nan M4 Pro Mac mini and M4 Pro/Max MacBook Pro. Developers won't request immoderate typical hardware to build clusters, conscionable modular Thunderbolt 5 cables and compatible Macs.

In a demo, I watched arsenic a cluster of 4 Mac Studios loaded and ran that monolithic Kimi-K2-Thinking exemplary successful an early type of ExoLabs's EXO 1.0. Notably, nan cluster utilized little than 500 watts of power, which is astir 10 times little than a emblematic GPU cluster (NVIDIA’s RTX 5090 is rated for 575W, but its demands tin besides jump higher).

macOS Tahoe 26.2 will besides springiness Apple’s unfastened root MLX task afloat entree to nan neural accelerators connected nan M5 chip, which should dramatically velocity up AI inferencing. Ironically, though, nan only M5 Mac disposable coming — nan 14-inch MacBook Pro — only supports Thunderbolt 4. That intends it won’t beryllium capable to return advantage of nan caller Mac clustering capability.

The unified representation and debased powerfulness creation of Apple Silicon already made Macs a useful prime for demanding AI work, but nan expertise to cluster aggregate systems together complete Thunderbolt 5 is perchance moreover much tempting to anyone moving pinch ample models. Of course, a Mac Studio pinch 512GB of RAM isn't inexpensive -- it starts astatine $9,499 pinch nan M3 Ultra spot -- but that's only nan highest-end option. Labs and companies that already person Mac Studios, Mac minis and MacBook Pros could perchance cluster systems they've already purchased.