Apple Mac Studio With M5 Ultra Review: There’s Simply Nothing Else Like It

Apple continues to cram more and more capable components into the same Mac Studio compact desktop design it debuted in 2022, and I keep waiting for it to hit peak power. With its flagship model powered by the M5 Ultra, it has reached a new peak — at least for now — and cemented its place at the front of the pack.

The Mac Studio gets expensive in mid- to high-end configurations with 128GB of RAM or more — typical workstation pricing, now with an added RAMaggedon cost premium! The model I tested, with the M5 Ultra, 256GB RAM and a 4TB SSD, runs about $12,300. It sounds like a lot because most people wouldn’t think of spending that much on a computer. 

But its price isn’t out of line by workstation standards, and there’s simply nothing else like the Mac Studio. All the compact competitors like DGX Stations, upcoming RTX Spark Windows systems and less powerful AMD Ryzen Halo or Intel Core Ultra-based models top out at 128GB. They’re more Mac Mini and low-end Studio competitors. And until I get a chance to test the Spark systems, I don’t know how their performance will compare, but their memory cap may be a constraint for some buyers.

What’s new for the 2026 Mac Studio

Aside from the refreshed chips and upgrades such as Wi-Fi 7 and Bluetooth 6 on the Ultra model — which can improve stability and reduce latency — the new Mac Studio models are physically identical.

Josh Goldman/CNET

There’s really just one new capability over the M3 Ultra and M4 generation systems: genlock over USB-C. Genlock is the ability to use an external, hardware-based signal generator, in this case connected via USB, to synchronize cameras for multicam capture and editing. It only works on the latest generation of M5-family desktops — M5 Pro Mac Mini and M5 Max and M5 Ultra Mac Studios, leaving the M6 Mini out as an inexpensive hub option.

Unlike the other versions of the chips, Apple’s top-end Ultra is on an every-other-generation upgrade cycle. So the last one was based on the M3 design, which has a GPU architecture that significantly lags the M5 for both graphics (especially ray tracing) and AI. 

The last generation of Ultra chips was more or less two Max versions linked by high-speed circuitry. Now, thanks to some architectural changes Apple introduced with the M5, the company could decouple the GPU scaling from the rest of the chip.

The M5 Ultra I tested is based on the M5 Max with 40 GPU and 18 CPU cores for a total of 80 GPU and 36 CPU (12 super and 24 performance) cores, plus a 32-core Neural Engine. That’s the same number of NE cores as the previous Ultra, but a more recent generation. But the lower-end Ultra is part low-end M5 Pro — double the CPU cores, with 30 — and part high-end, with 40 GPU cores instead of what would otherwise have been 32.

Like the other M5 chips, the Ultra incorporates on-GPU neural accelerators, which handle AI-intensive tensor (n-dimensional matrix) calculation. But Apple jacked up the memory bandwidth to 1.2TB per second, up from 813GB per second for the M3 Ultra and from 614GB per second for the M5 Max.

Configuration options

The configuration options are a little more flexible than they used to be, but you’re still hamstrung when it comes to memory. Apple’s not alone here, though, since nearly everyone now uses SoCs with on-package memory for all but desktop gaming systems.

The base $2,499 M5 Max configuration comes with only 36GB of memory, and that’s it — you can’t get more RAM without spending $600 to bump to the 18C CPU/40C GPU chip, which comes with 48GB of memory. Once you’ve moved up, you have a lot more leeway in memory configurations than with the last generation, up to 128GB.

The entry M5 Ultra model, with double the memory and double the storage, runs another $2,400. In fact, an M5 Max with 128GB memory (there’s no 96GB option) and 1TB storage is $100 less than the base Ultra configuration.

If you want 256GB, you have to jump to at least the base M5 Ultra chip. And for 512GB of RAM, you have to opt for the top-end chip. That option won’t be available until the end of October, though.

None of this would be a big problem if you could upgrade the memory in the Studio. On one hand, given high and rising component prices, it probably makes sense to spring for it now, since prices may never return to where they were pre-datacenter boom. But the flip side is that because of the high prices, the cost of memory now can be painful when you hit the triple digits. 

While 96GB as the base for the Ultra is pretty good, the next step up is 256GB for a whopping $4,000 bump. In contrast, the 128GB configuration of the M5 Max is an extra $2,000 over the base price, or $1,600 over the 64GB model. An in-between choice would be nice for the Ultra. All told, a maxxed-out configuration can run you a little over $18,000, at least until the 512GB option becomes available in October. We don’t know how much that will cost, but $8,000 over the base seems like a safe ballpark guess to me.

If you do think the 36GB of memory of the base model will last you for the next three-plus years, then you probably don’t need a Mac Studio and might be better off with a MacBook Pro M5 Pro or a kitted-out Mac Mini. The CPU and GPU in the Max and Ultra do provide better performance, but that potential speed boost could be throttled by insufficient memory.

It also might come down to how many USB ports you need; it’s annoying that you have to spend $5,499 for the base Ultra model just to get Thunderbolt ports (for 40 or 80Gbps USB-C data transfer) in the front instead of slow 10Gbps USB 3 (not even USB 4). My Mac Studios are always crammed into a relatively small space where it’s really awkward to get to the ports in the back, so I like to connect my removable drives to the front.

Performance analysis for the 2026 Mac Studio

Because more cores almost always deliver better performance, it’s unsurprising that the M5 Ultra-based system is thus far at the top of Apple’s heap across the board for GPU and multicore CPU speed. While the M6 in the Mini outdoes the M5 series, it’s only by about 6-8% (per Geekbench 7 and Cinebench 2026) on single-core performance. When you pile all those cores up in the Ultra, the CPU is damn fast.

The M6’s Neural Engine also outperforms the M5 Ultra by about 10% (per Procyon Computer Vision 2) for the Neural Engine. The latter handles everyday productivity tasks such as background replacement in videoconferencing, summarizing and drafting documents and so on. Think of that as saving a minute on a 10-minute task. The Ultra’s NE is faster than the Max’s in some ways, but the Ultra also has twice the NE cores.

Both the Max and Ultra are targeted at higher-end creative tasks and the development of generative AI tasks, where the high-core-count GPU pays off. For example, on the Procyon Stable Diffusion XL test, the 80-core Ultra scored 72% higher than the 40-core Max. And on Cinebench 2026’s rendering, the Ultra’s GPU scored about 46% higher than the Max. The closest a non-Mac came to the Ultra (based on our admittedly small sample) was the GeForce RTX 5090 running at full power in the Alienware 18 Area-51 gaming laptop: The Ultra scored 30% higher.

Josh Goldman/CNET

Apple pitches this as an AI workhorse — that’s really why the configurations for 256GB and 512GB exist — and like most of the high-powered systems these days intended for on-device processing, that means things such as agentic development, complex video generation, in-depth research, moderate-level machine learning and more. 

We’re all still looking for meaningful ways to benchmark system performance for these. But all these operations rely heavily on both CPU and GPU performance, and there the Ultra clearly has the most raw processing power in the Mac line. And the ability to cluster up to four Macs to pool memory and processing (as long as they’re equipped with Thunderbolt 5), a feature introduced with MacOS 26, makes the platform a pretty powerful option.

One benefit of the push for higher GPU core counts to serve AI is better gaming. The Max versus Ultra shows about a 20% increase in frame rates on Shadow of the Tomb Raider, but it also meant a step-up-ish in resolution: 150fps on the Max at 1,920×1,200, but 158 fps on the Ultra at 2,560×1,440.

Most notably, the M5 Ultra is the first Mac where I’ve seen passable frame rates using path tracing, at least in Cyberpunk 2077 (in the 60 fps range at 1080p). But that’s with MetalFX upscaling turned on, which looks awful; upscaling and raytracing frequently don’t mix very well. It performs really well overall, getting 118 fps at 4K for Shadow of the Tomb Raider, 177 fps at 1080p and 128 fps at 1440p for Cyberpunk, the latter with MetalFX turned off. 

But gaming is almost beside the point, because there are still only a handful of games available for the Mac that can take advantage of the GPU power, and I don’t think that will change any time soon.

Beyond the numbers, the system feels fast at everything, though I’ll probably have more to say about that once I use it some more.

I like almost everything about the Mac Studio hardware. It’s a class-leading compact system that, given you spend enough money on it, can do almost anything you want on a desktop. The other consideration is that if you need the power and 256GB-plus memory, the Mac Studio is the only Mac that will serve your needs; it’s just not an option in a MacBook Pro. 

You’ll know if you need the highest-end configurations because the size of the models you work with locally is easy to determine, and if you’re blowing your token budget on models and agents that you can bring on-device, it might be worth the investment (unless your company is bearing the cost). But unless you have unlimited funds, you may want to first figure out if less expensive configurations or other systems will serve your needs as well.

Performance charts

Note: Apple’s processor performance has historically been extremely consistent across devices, so I consider including MacBook Pro results as comparing apples to Apples.