In the latest macOS update, Apple has brought significant AI performance improvements to the M5 chip. However, the practical experience of this combination of punches has been greatly compromised due to the lag in the hardware interface.
Native support for neural accelerators
The core part of this macOS 26.2 Tahoe update is to enable the open source machine learning framework MLX to directly call the dedicated neural accelerator in the M5 chip GPU core. In the past, developers had to use underlying tools such as Metal or Core ML to get involved with this part of the hardware, and the whole process was relatively complicated.
At this time, the native support of the MLX framework simplifies the process, allowing developers to operate more conveniently with these hardware units. This shows that when conducting AI model training or inference on a single M5 Mac, computing efficiency will be truly improved, especially for large language model applications on the device.
Specific manifestations of performance improvement
Based on official data from Apple, when the large language model is run and the initial prompt is processed, the peak AI performance of the M5 chip can reach the highest level, which is four times that of the previous generation M4 chip. This improvement is mainly the result of the full participation of the neural accelerator.
Obtaining stronger computing power on a local Mac means that individuals or small teams engaged in machine learning research do not need to rely on remote servers. This lowers the threshold for AI development and makes prototype verification and iteration faster.
Introducing Thunderbolt 5 cluster computing
Another highlight of this update is the introduction of cluster computing support for the Thunderbolt 5 interface. Researchers can use Thunderbolt cables to connect multiple Macs together to build a computing cluster that works together to process a single large AI model.
Thunderbolt 5 has a bandwidth of up to 120Gbps, which far exceeds the 40Gbps bandwidth of Thunderbolt 4. This higher bandwidth allows data exchange between multiple devices to become faster, thereby clearing key obstacles for distributed computing.
Technological Innovation in Cluster Computing
In addition to doubling the bandwidth, Thunderbolt 5 also has the ability to assist remote direct memory access technology. This special technology can give one Mac in the cluster the right to directly read and write the memory contained in another Mac without the need to relay it through the CPU. Based on this, the delay caused by data transmission can be greatly reduced.
This shows that the internal communication efficiency of a computing cluster composed of multiple Macs is close to the internal level of a single machine. Such a capability is crucial for large model training tasks that require frequent exchange of intermediate calculation results.
Realistic constraints of current hardware
However, an actual contradiction is that the M5 Macs currently sold on the market, such as the 14-inch MacBook Pro, are still equipped with Thunderbolt 4 interfaces. This makes it impossible for users to enjoy all the advantages of the two updates on their existing devices at the same time.
Users are faced with a choice, either to use the neural accelerator on a single M5 Mac to obtain a huge improvement in local performance, or to use multiple older Macs to build a cluster through Thunderbolt 5. There is no way to have both at the same time.
The full potential of future devices
Waiting for the new Mac device to come out, the device must integrate both the M5 chip and the Thunderbolt 5 interface to fully unlock the full potential of this system update. It is expected that high-end models such as Mac Studio or Mac Pro will meet this condition in the future.
By then, researchers will be able to set up a Mac computing cluster, which has both powerful single computing power and ultra-high-speed interconnect capabilities, which will effectively transform Apple's desktop devices into a powerful, integrated AI research and development platform.
In your opinion, would it be wiser to wait for the new M5 Mac equipped with Thunderbolt 5, or would it be more pragmatic to start research projects with existing equipment now? Welcome to share your personal opinions in the comment area. If you feel that this article is helpful, please give it a like to support it.
