Apple Mac Studio With M4 Max Review: Big Gains, Petite Package

We've kind of struggled in the past with the right machine to compare against the Mac Studio. Tiny, tiny desktops are needlessly small, while big desktops can use way more power and as a result offer a lot more performance. And then AMD designed and released its Ryzen AI Max processors built upon its Strix Halo chip. It's a monolithic design with a huge Zen 5 CPU up to 16 cores and the biggest integrated GPU we've ever seen in an x86 processor, in the Radeon 8060S.

The only retail product with this family of CPUs right now the ASUS ROG Flow z13, which is a convertible laptop that we reviewed recently. Given its similar price and thermal properties, it's a great point of comparison. It's worth noting that the Ryzen AI Max+ 395 can stretch its legs up to 120 Watts for its TDP, but ASUS has limited it to 80 Watts in its tablet form factor. 


So with that out of the way, it's time to dive in. Some of our past Macs are back for more comparisons, including an M2 Pro MacBook Pro 14" from early 2023 and an M3 Max MacBook Pro 16" that we haven't reviewed, but is available to us on loan. We don't have full benchmarks for that machine, but we have enough to get an idea of performance. We've also got the full-fat M4 Pro inside a 14" MacBook Pro (I would have loved to review this last fall, but time was working against me) and have also been fortunate enough to borrow an M4 Pro Mac mini, which is the base model of that configuration with a binned M4 Pro for $1399. 


So with the full lineup detailed, let's take a look at performance. All of the Macs are running the latest version of macOS Sequoia 15.3.2 as of this writing, while the ROG Flow z13 has a fully updated Windows 11 installation. Discrete drivers for Apple hardware aren't really a thing as they're distributed with the operating system itself, so having the latest OS also means having up to date drivers too. 

Mac Studio M4 Max Disk Benchmarks

SSD testing on macOS is still possible thanks to ATTO Disk Benchmark and AmorphousDiskMark, the latter of which is more-or-less equivalent to Crystal Disk Mark. The default settings wind up being more of a cache test than anything, so we set the write pattern to Random and set the file size to 32GB. 




Apple's SSDs in its desktops are on discrete cards these days, but they're just cards with brainless NAND modules. The NVMe controller is built into the M4 Max SoC itself. That controller and the 1TB of NAND are monsters, however, as they routinely eclipse 6 GB/sec in reads and over 7 GB/sec in writes. That seems backwards, or maybe like there's some cache involved, but as expected Apple's black box is quite opaque. 

Speedometer 3 Web Benchmark

We use BrowserBench.org's Speedometer test to take a holistic look at web application performance. This test automatically loads and runs a variety of sample web apps using the most popular web development frameworks around, including React, Angular, Ember.js, and more. This test is a better example of how systems cope with real web applications in comparison with a more compute-focused test like JetStream. All tests were performed using the latest version of Chrome, which is version 134 at the time of this writing.


Apple Silicon is just crazy fast at web-related tasks. The difference between any of the M4-based Macs and the ASUS ROG Flow z13 had us going back and re-running the test. The latest Chrome did goose it by a couple of points, but it still loses out to the lowly (and two-years-old) M2 Pro MacBook Pro 14. 

Cinebench 2024 Rendering Benchmark

Next up is the latest-generation 3D rendering benchmark from Maxon, based on the Cinema 4D rendering engine. It's a purely CPU-based test that doesn't make use of the graphics processor at all, and it scales very well with additional CPU cores. We ran both single- and multi-threaded tests on all of the machines in the charts.


The Strix Halo-powered ASUS machine might have more cores and many more threads (32 with SMT) over all, but the Mac Studio's single-threaded performance is high enough and multi-core scaling is strong enough to let it claim the top spot here. All of our M4-based systems have a significant single-threaded advantage here, so it's not just a matter of memory bandwidth at play. In what is sure to be a common thread among our results, the MacBook Pro 14" with M2 Pro settles into last place by a rather wide margin, highlighting just how much Apple has improved its CPUs over the last couple of years. 

Geekbench 6 System Benchmarks

Geekbench is a cross-platform benchmark that simulates real-world workloads in a wide variety of tasks, including encryption, image processing, physical simulation, machine learning, and many more. We tested the systems featured here with the latest Geekbench 6 version to get an idea of their overall system performance.


This benchmark has a history of favoring Apple pretty heavily. And again, the M4 Macs all flex their single-threaded prowess. Despite the full-fat Strix Halo chip in the Asus, the Mac Studio managed to wrestle the top multi-threaded spot from the Ryzen AI Max+ 395 in the ROG Flow z13. Even the full 14-core Mac Pro managed to take a win over the Asus. 

UL Procyon Office Benchmarks

While we often run PCMark 10's test suite, it's not available for macOS. However, Prycon is available and uses Microsoft Office products to perform its own cross-platform test of productivity workloads. Let's see how these systems fare. 


There's definitely a gathering of leaders here. It's lightly-threaded enough that the Mac mini's binned M4 Pro losing two performance cores doesn't substantially lag behind the full-fat version in the MacBook Pro. There's no real clear winner though; the Mac Studio takes the top spot by less than two percent, which we chalk up to a margin of error. 

UL Procyon Computer Vision Machine Learning Benchmarks

The idea of "edge AI", or running AI workloads natively on your local devices instead of in the cloud, is only just emerging on mainstream PCs. As such, benchmarks for these workloads aren't exactly prolific. Fortunately, UL has already built a few into its Procyon benchmark suite. The following is a look at how some of our machines do in this benchmark suite's AI Computer Vision benchmark. This test exercises the subject's ability to handle machine vision workloads, which you'll find in everyday tasks like webcam background blur, subject tracking, and eye gaze correction, for a few examples.


Apple makes a pretty big deal about its AI -- that's Apple Intelligence -- but honestly the performance is pretty pedestrian. The ASUS ROG Flow z13 takes the top spot in both the NPU and GPU benchmarks. These were done with integer data types, which is the most common for real-world models among the data types available for testing. The Mac Studio's NPU is definitely faster than its predecessors again, but eats AMD's dust. 

Geekbench AI Machine Learning Benchmarks

The Geekbench AI benchmark provides a straightforward look at how well a device handles a variety of AI-assisted tasks. This quick and easy test gives you a numerical snapshot of the CPU, GPU, and NPU’s ability to power through real-world machine learning workloads, factoring in both speed and accuracy. The higher the score, the better the device’s AI chops, whether it’s image recognition, object detection, or natural language processing. Results are presented in three levels of numerical precision: single precision or FP32, half precision or FP16, and quantized or INT8, the latter of which is a very common precision for on-device edge AI workloads. All results that the benchmark provides are geomean scores from multiple runs of each test workload.


The ROG Flow z13 claims a victory in both CPU and GPU workloads on behalf of ASUS and AMD. While the Mac Studio is the fastest Mac with its M4 Max SoC, its CPU and GPU are not enough to come close to overtaking the lead. Geekbench AI has not been updated to work with AMD's NPU just yet, and so that chart is all Macs, and we can see that the M4 generation represents a pretty substantial increase of around 50% compared to the M3 and earlier generations in both half-precision and quantized data types. 

Next up, we'll take a look at some graphically-intensive tests...