A video blogger showcased a computing cluster made of new Apple Mac mini units featuring the M4 processor, demonstrating that it can sometimes outperform a powerful graphics card.
Many believe that acquiring a more powerful computer means purchasing one expensive device. However, there are alternative methods to perform extensive computation. The concept of clusters allows the use of multiple computers, or at least computing units. Working together for parallel task execution significantly reduces processing time.
In a YouTube video, enthusiast Alex Ziskind displays the setup for clustered computing using five M4 Mac minis. The cluster receives tasks that are distributed across all machines. Typical small clusters rely on Ethernet for communication between nodes, but the YouTuber took advantage of Thunderbolt connectivity via Thunderbolt Bridge. This significantly accelerates communication between nodes and enables larger data packet transfers.
Ethernet can operate at 1 Gbps under normal conditions or up to 10 Gbps when equipped with compatible computers. In contrast, Thunderbolt Bridge achieves speeds of up to 40 Gbps for Thunderbolt 4 ports or 80 Gbps for Thunderbolt 5 in bidirectional mode on models with M4 Pro and M4 Max chips.
Ziskind notes that using Apple Silicon for clustered computing could be more economical than a PC with a high-end graphics card. GPU data processing relies heavily on the availability of substantial video memory. For instance, a graphics card may have 8 GB, which is not much, even for gaming. The use of unified memory on Apple Silicon imposes fewer constraints on configurations and allows for larger memory pools—essentially, the Apple Silicon GPU has access to significantly more memory, especially in cases like the Mac with 32 GB of RAM.
Additionally, graphics cards consume a lot of power. This high energy consumption results in increased ongoing operational expenses. It turns out that a cluster of five Mac minis consumes less power than a single high-performance graphics card.
To set up the cluster, Alex Ziskind uses MLX, Apple's open-source project described as "an array framework designed for efficient and flexible machine learning research on Apple Silicon." MLX operates using the standard distributed computing methodology, MPI. The project can run multiple Mac computers of varying performance without significant hardware expenses. Among other features, MLX is optimized for small clusters.
Effective, but Not Always
While combining the performance of several Mac minis into a cluster seems appealing, this approach does not benefit every task. There are virtually no advantages for regular Mac usage—such as running applications or games. The technology is geared towards processing large datasets or high-intensity tasks that benefit from parallel processing. This makes the cluster ideal for AI applications, particularly language models (LLM).
Moreover, it is not the simplest computing solution for the average Mac user. During tests, Ziskind found that purchasing a Mac with M4 Pro delivers better LLM performance than two M4 computers in a cluster. Such a cluster may be useful when greater performance is needed than a single powerful Mac can provide. If a model is too large to run on one Mac, for instance due to memory limitations, a cluster can offer more capacity.
The enthusiast claims that at this stage, a high-end Mac with the M4 Max and a large memory capacity is more efficient than a cluster of less powerful machines. However, if the task requirements somehow exceed the capabilities of the highest Mac configuration, the cluster may provide assistance.
Nonetheless, there are still some limitations to consider. The experimenter had to resort to using a Thunderbolt hub to connect the nodes to the host, which reduced the available bandwidth. Directly connecting the computers resolved the issue, but this method limits the number of available ports and scalability. Additionally, temperature must be monitored—the primary Mac mini in the configuration was extremely hot. Alex Ziskind acknowledged the Mac mini cluster tower experiment as interesting but does not plan on using it long-term.
Source: Apple Insider
Comments (0)
There are no comments for now