Apple Mac Studio With M5 Max: First Impressions

We check out the M5 Max-equipped model The same, but different. The new Mac Studio models feature M5 Max and M5 Ultra silicon, dramatically improving the performance of audio-based AI workloads.

By Mark Wherry

Back at the end of August, Apple unveiled new Mac Studio models that begin shipping this week. Unlike the previous Mac Studio generation, where different M-series silicon was used in the standard and high-end configurations - an M4 Max or M3 Ultra respectively - the new machines all boast M5-series silicon. The standard model features an M5 Max, which debuted earlier in the year with the latest MacBook Pros, while the high-end model is powered by the brand new M5 Ultra.

The release of the new Mac Studio coincides with the recent launch of macOS 27, affectionately known as Golden Gate, which comes pre-installed - although you might be prompted to download an upgrade during the setup process. From a performance perspective, macOS 27 seems very promising, although there are the usual compatibility caveats musicians and audio engineers can expect with every major new operating system upgrade. As always, it's best to consult any support information published by the developers of any applications or plug-ins on which you rely before committing yourself to a macOS 27-based system.

As you might expect, running Logic Pro on a new Mac Studio with macOS 27 seemed to work just fine in my initial testing. The Mac Studio in question was equipped with an M5 Max (featuring 18 CPU cores - comprising six super and 12 performance cores - 40 GPU cores, and 128GB unified memory). The standard model's M5 Max is equipped with the same CPU configuration but includes a 32-core GPU and 36GB unified memory.

In terms of audio performance, the Mac Studio with M5 Max offered a 6 percent improvement over an M4 Max-based MacBook Pro (also with 128GB unified memory). With the Processing Threads set to Automatic, the new Mac Studio played back 604 stereo tracks (each equipped with two Amp Designer instances), while it was possible to increase this limit to 660 tracks by selecting the maximum setting (utilising all 18 cores). On the MacBook Pro with M4 Max the track counts were 572 and 622 with Processing Threads set to Automatic and the maximum setting (using the 16 available cores) respectively.

Although these improvements aren't significant in and of themselves, it should be remembered they reflect just one aspect of the M5's architecture: CPU performance. Because while it used to be the case that the GPU was less important for music and audio applications, beyond having sufficient resources for Metal to accelerate rendering the user interface, this is no longer the case. As modern music and audio software - which includes both applications and plug-ins - embrace more AI-based functionality, other compute domains become dramatically more important.

One of Apple's intentions for the new Mac Studio, whether configured with an M5 Max or Ultra, is to be the ultimate desktop for on-device AI and the world's most demanding pro workflows according to Apple's Chief Hardware Officer, Johny Srouji. And one of the many ways Apple hopes to achieve this ambition can be found in the next-generation GPU architecture the company engineered for the M5, which includes Neural Accelerators.

These Neural Accelerators could almost have been called matrix accelerators' because, as you may know, AI compute relies heavily on matrix multiplication. And one reason GPUs have been ideally suited to this kind of work, aside from their ability to perform large numbers of calculations in parallel, is that GPU cores can perform matrix multiplication using their general-purpose arithmetic hardware. However, to improve AI performance further, Apple added a dedicated matrix-multiply execution block to each GPU shader core, which is conceptually similar to what Nvidia calls a Tensor Core. So, rather than performing all matrix calculations using the GPU's general-purpose arithmetic pipelines, suitable matrix-multiplication operations can instead be executed by specialised hardware that dramatically increases their throughput.

Now, if you're wondering what this has to do with audio and music, audio-based AI workloads often involve applying large numbers of floating-point calculations to blocks of sample or spectral data, making them particularly well suited to the highly parallel processing offered by a GPU. And since the neural networks behind these workloads make extensive use of matrix multiplication, the Neural Accelerators turn out to be really useful for accelerating AI-based audio processing.

To see how this translates into a real-world, if slightly over-the-top, AI-based audio workload, I turned to Logic Pro's Stem Splitter feature. For my test material, I bounced the stereo, 24-bit/96kHz AIFF download version of Peter Gabriel's album Us into a single Wave file, which resulted in a 2GB file containing nearly an hour's worth of audio. I then dragged this back into a new Logic Project (running at 96kHz) and used the Stem Splitter command. While it was running, the resulting GPU activity was clearly visible using the Terminal utility macmon with the CPU barely being touched.

This screenshot shows the Stem Splitter command running in Logic Pro on a new Mac Studio with M5 Max. Notice how macmon (running in the Terminal window) shows almost no CPU usage while the GPU cores are fully utilised.

Using the new Mac Studio with M5 Max, it took 108 seconds to generate the six stems - and it did a pretty good job. However, to understand how impressive this result is, I repeated the same test with the MacBook Pro with M4 Max; this time, Stem Splitter took 232 seconds to complete the same test approximately 2.16x slower. Testing with Apple's previous high-end Mac Studio, powered by an M3 U

LINK:	https://www.soundonsound.com/news/apple-mac-studio-m5-max-first-impres...
	See more stories from soundonsound

Most recent headlines