In addition to importing in macOS 26.2Edge Light featureThis allows the MacBook screen to instantly become a video flashlight, and Apple has also added a feature in this version that allows multiple Macs to be connected together to create an "AI supercomputer".
In the upcoming macOS 26.2 update, Apple is adding a new low-latency feature that allows users to connect multiple Macs together via Thunderbolt 5 cables. For developers and researchers, this offers a potentially efficient way to build powerful AI supercomputers from existing Mac devices, enabling them to run large local models.
Unlocking 80Gb/s of transmission bandwidth through Thunderbolt 5, running a 1M parameter model
While there have been previous attempts to cluster Macs, these were often limited by slow transfer speeds (sometimes down to 10Gb/s). The new features in macOS 26.2, however, fully unleash Thunderbolt 5's connectivity capabilities, reaching up to 80Gb/s.
This cluster feature is not limited to the expensive Mac Studio; it also supports the Mac mini with the M4 Pro processor, as well as the MacBook Pro with the M4 Pro and M4 Max processors. Developers do not need special hardware; they only need a standard Thunderbolt 5 cable and a compatible Mac model to build it.
In the demonstration, a computing cluster consisting of four Mac Studio machines (each with up to 512GB of unified memory) was able to successfully load and run the Kimi-K2-Thinking model with 1 trillion parameters.
With a power consumption of only 500W, its energy efficiency is significantly better than that of GPU clusters.
It is worth noting that this computing cluster, consisting of four Mac Studio machines, consumes less than 500 watts of power when running large models.
This data is astonishing compared to traditional PC GPU solutions. For comparison, a single NVIDIA RTX 5090 graphics card has a rated power consumption exceeding 575W (actual demand may be even higher). In other words, the Mac cluster consumes approximately one-tenth the power of a typical GPU cluster, demonstrating a significant advantage in energy efficiency.
MLX supports the M5 neural accelerator, but hardware support is "awkward".
In addition, macOS 26.2 will grant Apple full access to the open-source MLX project, allowing it to directly access the Neural Accelerator on the M5 processor, which will significantly improve the speed of AI inference.
However, there is a hardware limitation: the only Mac model currently on the market with an M5 processor—the 14-inch MacBook Pro—only supports Thunderbolt 4 ports, meaning that this latest M5 processor laptop cannot yet utilize the latest Thunderbolt 5 clustering feature.
While the top-of-the-line Mac Studio (M3 Ultra processor, 512GB memory) is expensive (starting from $9499), this feature offers an attractive option for labs and businesses that already own a Mac Studio, Mac mini, or MacBook Pro, allowing them to cascade their existing devices into powerful tools for processing large AI models.
Macs originally had computing stacking capabilities.
In fact, Apple has long enabled Macs to have computing stacking capabilities; for example, during WWDC 2019, they demonstrated the use of multiple Mac minis connected in series.Construct a collaborative computing serverThrough computing performance virtualization, devices such as the iPad Pro can also be adapted to higher computing performance execution modes, while also serving as cloud servers, and the number of Mac minis that can be connected can be increased as needed.
AWS also showcased [its products/services] during the re:Invent 2020 conference.EC2 Mac Executable Unit Based on Mac miniThis allows developers of iOS and Mac apps to build an environment suitable for Apple's operating system services in a shorter time and to expand storage capacity according to traffic usage needs. It is also used with AWS pay-as-you-go pricing.
