Apple M4 & M3 Ultra Macs

A new M in charge: will the M3 Ultra achieve pro audio domination?

Apple’s M‑series Macs continue to power ahead. We benchmark the latest Mac mini, MacBook Pro and Mac Studio models for audio performance.

With one exception, Apple have upgraded or reintroduced new models of every single Macintosh in the last few months. Last October, the company unveiled updates to their iMac and MacBook Pro families — in addition to a brand‑new Mac mini — with March bringing an updated MacBook Air and Mac Studio, which takes the crown as “the most powerful Mac ever”. Most of these are based around their latest M4 processor. The M4 debuted in the iPad Pro back in May 2024, but it was only a matter of time before Apple brought this latest silicon to the Mac, and, in doing so, continued the established pattern of augmenting the standard chip configuration with more powerful Pro and Max variants.

What’s It All M4?

The M4 represents a significant leap over the M1 and M2 architectures already familiar to musicians and audio engineers, building on the M3 chips released a year earlier. The M3 family were the first personal computer chips to be fabricated on a 3nm process node using TSMC’s N3 technology. This was significant because, simply put, moving to a smaller process allows for tinier transistors and greater transistor density, which can afford improvements in both performance and power efficiency. Prior M‑series chips were fabricated on a 5nm process, so even if Apple had just manufactured the existing M2 chip with the newer N3 technology, it would theoretically have resulted in a more energy‑efficient chip with better performance.

However, the M3 was more than just the sum of its process. Where the M2 had been an evolution and refinement of the M1, offering around a 20 percent increase in real‑time audio performance, the M3 represented a much bigger leap. Thanks to significant advances in both manufacturing and architecture, the M3 Max delivered 58 to 75 percent more audio processing compared to the M2 Max, depending on the host application. So key were the M3’s gains that the M3 Ultra can outperform an M4 Max — much as an M1 Ultra could surpass an M2 Max. Consequently, whereas the base model in the new Mac Studio line features an M4 Max CPU, the top‑end model is based around the new M3 Ultra.

Core Performance

As you may already be aware, M‑series chips employ a heterogeneous architecture for general‑purpose computing tasks. This means that while every CPU core supports the same basic instruction set capable of executing the same code, there are two types of cores, with different performance and efficiency characteristics. Performance (P) CPU cores are finely tuned for high performance, whereas Efficiency (E) cores are optimised for handling less demanding tasks using the least amount of energy. Music applications typically rely on P cores for real‑time audio, leaving the E cores to handle other, non‑real‑time tasks such as lookahead processing.

However, applications can’t directly assign specific tasks to certain cores. Instead, macOS 11 introduced the Audio Workgroups API, enabling applications or plug‑ins to flag certain tasks as being responsible for handling real‑time audio. This, in turn, allows macOS to optimise the usage of performance and efficiency cores to achieve glitch‑free playback. And although such a scheduling system isn’t perfect, it generally provides the best possible overall performance. Because whilst one application or plug‑in can’t know the specific processing requirements of another, macOS knows everything — at least in terms of an application’s needs.

The M3 was a notable advancement over the M2, introducing a next‑generation GPU architecture alongside many key advancements to the design of the CPU cores, such as improved branch prediction. Simply put, branches are points in a program where a condition needs to be evaluated to figure out what happens next. A naïve example might be to check if a channel is muted on a mixer: if the channel is muted, the program can skip over that channel and proceed to the next one, but if the channel is not muted, the program will need to process the playback required from that channel.

To speed up the execution of our modest mixer, a branch predictor would make an educated guess as to whether a given channel is muted or not based on its previous state. If the branch predictor guesses correctly, the CPU will have already started processing the correct instructions ahead of time, allowing the program to run smoothly. However, if the prediction is wrong, the CPU must discard the instructions that were speculatively processed and reprocess the correct ones, which can adversely affect performance.

In practice, branch prediction works at a much lower level, enhancing DSP algorithms where decision making is inherently part of the process. For example, a limiter needs to decide whether incoming audio samples are above or below the threshold. If the current sample is above the threshold, it’s very likely this condition will be true for the next sample. Therefore, the CPU can predict the outcome for the next cycle.

The P cores also feature wider decode and execution engines, both of which enhance audio processing performance. A wider decode engine allows more instructions to be processed simultaneously, while a wider execution engine enables more of those instructions to be executed in parallel. Such improvements benefit both traditional DSP algorithms — especially computationally intensive effects like convolution reverb — and modern machine‑learning‑based audio effects for restoration and assisted mixing.

Apple claimed the M3’s performance cores were up to 30 and 15 percent faster compared to the M1 and M2 respectively, with the efficiency cores achieving a greater increase of up to 50 and 30 percent. Perhaps most impressively, though, and thanks in no small part to the 3nm process, the M3’s architecture could deliver the same multi‑threaded CPU performance as the M1 using half the power.

From London To South Wales

The M4 builds on these architectural advancements, with the performance core offering further improvements in branch prediction, the ability to decode 10 instructions simultaneously, a 40 percent larger reorder buffer, and what Apple refer to as “Next Generation ML [Machine Learning] Accelerators” (see box). The larger reorder buffer helps to improve general compute performance, but has fewer direct benefits when running DSP code, for reasons that are a little beyond the scope of this article. However, it can still improve the performance of audio applications and plug‑ins, for example, by speeding up the execution of user interface handling and so leaving more resources available for executing DSP code.

Differ‑M‑tiation

The Pro and Max variants of the M1 and M2 families shared many similarities, particularly in terms of CPU core configurations. While the standard chips each featured four performance cores and four efficiency cores, the Pro and Max versions were equipped with eight P cores, with a more affordable Pro option offering six P cores. The M1 Pro and Max chips had two E cores, while the M2 Pro and Max chips offered four.

For the ultimate performance, Apple then introduced an Ultra variant. The M1 Ultra debuted with the Mac Studio, while the M2 Ultra arrived alongside the first Mac Pro based on Apple Silicon. However, unlike other M‑series chips, the Ultra combines two Max dies interconnected via an Apple‑designed high‑speed fabric known as UltraFusion. This technology allows the system to treat the package as a single unified chip with double the resources. As a result, the Ultra delivers not only twice the number of CPU and GPU cores, but also dual Neural and Media Engines.

The Pro and Max variants of the M1 and M2‑series chips were both available in identical 10‑ or 12‑core configurations, and consequently, systems containing these chips achieved similar levels of performance. However, the Pro and Max variants were still distinct in other ways such as the amounts of physical memory supported and the number of GPU cores and video encode and decode engines.

With the release of the M3 family, Apple kept the same eight‑core composition for the standard M3 chip, but decided to more significantly differentiate the number and distribution of cores in the Pro and Max configurations. The M3 Pro was available with either 11 or 12 CPU cores, comprising six E cores and either five or six P cores, whereas the M3 Max became further maximised by offering 14 or 16 CPU cores, featuring four E cores with either 10 or 12 P cores. The belated M3 Ultra combines two M3 Max dies, but unlike previous generations, is available in two variations, depending on which variant of the M3 Max it uses. Both offer eight E cores, and you can choose either 20 or 24 P cores.

With the M4 series, Apple have once again decided to rebalance the CPU core configurations. The standard M4 in the Mac is a 10‑core chip with six P cores and four E cores. The M4 Pro is available in 12‑ or 14‑core versions featuring four E cores and either eight or 10 P cores, while the M4 Max configurations have also been upgraded to comprise 12 or 14 cores, each featuring the same four E cores, adding either eight or 10 P cores. Another significant change with the standard M4 chip is the inclusion of 16GB unified memory as standard — arguably a long‑overdue decision. Previously, 8GB had been the minimum for base models.

As we will see, the real‑world difference in application performance between the M4 Max and M3 Ultra is, frankly, chasm‑like!

Figure 1 shows that the Geekbench results for Apple’s new chips are mostly as one would expect, aside from one noticeable anomaly: the multi‑core score for the new M3 Ultra chip. According to this result, the Mac Studio with M3 Ultra appears to be less than six percent more powerful than the M4 Max‑based model — an increase that would hardly justify the price gap, even considering other differences in the specifications. However, I think this result is indicative of a flaw in Geekbench’s benchmarking rather than being reflective of the M3 Ultra’s raw horsepower. As we will see, the real‑world difference in application performance between the M4 Max and M3 Ultra is, frankly, chasm‑like!

Figure 1: This chart shows the results of Geekbench 6’s single‑ and multi‑core CPU tests for the Macs featured in this review, along with results for older systems and the M4‑based iPad Pro. (The number of CPU cores are indicated in brackets for systems available in different configurations.)

The Mini Studio

The design of the new Mac mini with M4 clearly owes much to the Mac Studio; indeed, a more apropos name might have been the Mac Studio mini. Apple have shrunk the 7.7 x 3.7‑inch dimensions of the Mac Studio to a 5 x 2‑inch cuboid. An honourable mention must be given to the team that engineers Apple’s power supplies. It’s remarkable when you consider that, in addition to housing such a powerful system within a compact aluminium enclosure, Apple also manage to integrate a 155W power supply within the Mac mini’s tiny form factor. The Mac mini by itself could easily be mistaken for the power brick of a competing product!

Despite its size, the Mac mini is surprisingly capable!

Considering its size, the new Mac mini offers a surprising array of connectivity. The 3.5mm headphone jack has finally been moved to the front of the device, accompanied by two USB‑C ports with support for USB 3 (up to 10Gb/s). Sadly, however, this means there is no space for any USB‑A ports.

The power button, previously located at the rear of the Mac mini, has also been relocated. However, rather than moving this seemingly vital control to, say, the front, or maybe the side, Apple’s engineers have instead decided to place it underneath. This makes it hard for the button to be pressed accidentally, but, for the same reason, it also makes it hard for the button to be pressed intentionally. In practice, this is only really inconvenient during Touch ID setup, where a physical button press is required to confirm secure intent and establish a direct connection to the Secure Enclave.

A more important consideration is that the base model Mac mini with M4 and a 256GB SSD delivers lower data transfer rates when compared with the more expensive, higher‑capacity alternatives (as you can see from the Blackmagic Disk Speed Test results)....