Review: M5 Ultra Mac Studio: Pure, unadulterated power
Computerworld ·
I’m old enough to remember when Macs were friendly little machines. While a little underpowered and equipped with some platform-unique foibles, they were really good at some things, highly secure, and had the best user interface of any PC. That was then; this is now. And while the platform is still unique, still highly secure, and still has the best user interface, Apple’s all-new Mac Studio is a beast of a machine. It’s incredibly powerful and would crackle with energy if it weren’t so energy efficient. TL;DR review No matter what pro workflow you follow, the M5 Ultra Mac Studio is more than you need. Longer review Apple hasn’t changed the design of the Mac Studio. It’s 7.7-inch square, 3.7-inch high aluminum box with two USB-C and one SDXC port on the front, and an array of ports — four Thunderbolt 5, two USB A, 10Gb Ethernet, HDMI, and headphone — at rear. It brings Wi-Fi 7 and Bluetooth 6, supports up to eight displays with up to 6K resolution at 60Hz, or four Studio Display XDR units. Basically, this one box can power a bank of displays, and for those using it in a live video production environment, Apple has also introduced generator locking (a.k.a. genlock) over USB-C, which lets the on-display image synchronize precisely with a pro video camera for the most precise editing and playback you’ll find. Pros and cons Pros Massive CPU/GPU power capable of handling the most demanding 3D and AI tasks. Neural accelerators and unified memory make it ideal for training and running large LLMs on-premises privately. Doubled read/write performance over the previous generation via a new storage controller and fast NAND. Inclusion of four ProRes accelerators (handling up to 33 streams of 8K) and genlock over USB-C for precise sync. High performance at low wattage; runs quiet and cool compared to high-end PC alternatives. Features Thunderbolt 5, Wi-Fi 7, and Bluetooth 6. Works exceptionally well as a headless AI or rendering server. Cons Extremely high retail price ($12,299 for the top spec). Neither the RAM nor the SSD can be upgraded or replaced by the user. No keyboard or mouse included, which feels odd for a five-figure machine. For my review, Apple loaned me a top-of-the-line M5 Ultra Mac Studio equipped with a 36-core CPU and 80-core GPU. It has 256GB memory, 4TB storage, and carries a retail price of $12,299. At that price and with those specifications, it is of little surprise that this is such a great computer. It’s also intended for some of the most demanding 3D and AI tasks you’ll find out there, rather than for people who spend their time writing about tech — though this machine can open multiple tabs in Chrome without stuttering, so unleashes real productivity boosts even at my level. Apple Let’s get one thing out the way That’s not to say some pro users lack reservations about it. Many are frustrated that neither the RAM nor SSD are user serviceable, while the lack of a keyboard and mouse in the box chimes oddly in a computer that costs five figures. In Apple’s defense, in these pro markets most users probably already own better keyboards and mice than you’d find in the box, while the massive performance benefits you’ll find with the M5 Ultra chip are in part realized because RAM in Macs is built to be part of the processor itself. That SOC design boosts efficiency, cuts heat dissipation, and makes for great performance at low wattage relative to other machines. Changing the way memory is installed on the chip would compromise Apple’s silicon architecture, which Apple doesn’t want to do. It’s all about the architecture The Mac is powered by a quad-die chip built using Apple’s UltraFusion process. Inside is an engineering marvel. The 80-core GPU brings neural accelerators, making these Macs truly phenomenal machines for building, developing, training, or running AI models. Apple’s unified memory tech is a godsend for tasks like these as it scales to handle… well, most anything you throw at it. Efficiently. Apple’s coupled chip speed with performance across the system. That means an advanced storage controller, superfast NAND, and really fast, next-gen SSDs for twice the read/write performance of the last model. This basically means the chip can grab big dollops of data and very swiftly push them around the system all the way from storage to display. Then there’s the rest of the advanced tech to consider, including the inclusion of four ProRes accelerators, which means my test Mac can deal with up to an incredible 33 streams of 8K ProRes video. This is a machine that can happily handle massive multi-camera editing, racing through dozens of high-res angles at once to field the perfect shot. Think live sports, concerts, and movie shoots. Not only this, but all those streams are at normal res, no conversion required — no sitting around waiting for low-quality proxy files. Apple A bicycle for several minds I’m thinking a full-flight video rendering data server in an 8-inch box that consumes perhaps 10 cents an hour at peak power (c. 480 watts). I’m also thinking of it as an on-premises AI system for me, the family, or any enterprise. Compared to the previous equivalent mode, the M3 Ultra, Apple says the chip brings up to 4.3x faster AI performance, up to 1.8x faster GPU performance, and up to 1.3x faster CPU performance. It’s also almost ten times faster than the M1 Ultra for AI. None of this is accidental. All of it is designed. This whole creation is architectural; it leans heavily into Apple’s software and hardware integration, which now also extends to the design of the processor itself. Making the RAM user-serviceable would limit the performance of the machine, which at this price and in this sector of the market seems a little counterintuitive. Sure, you can build yourself something pretty powerful using a Ryzen 9 chip that consumes 900 watts at peak and runs hot. Or you can put Apple’s silver box on your desk and barely hear a thing as it crunches through ‘god tier’ AI models your Ryzen can’t handle without additional GPU’s. All the same, if you want to configure your own memory you do have a choice — it’s just not a Mac. Choice is nice I know what I’d choose. Based on a weekend of using the Studio, I’ve found what it does is beyond most of the feeble tests mere mortals like me can cook up, so I thought you might want some benchmarks: Geekbench 7 Single-core CPU : 3,771 Multi-core CPU : 52,350 GPU (Metal): 360,019 GPU (OpenCL): 214,466 Cinebench CPU (Multi-thread): 18,052 GPU: 141,480 Putting these numbers into context, Apple explains what these numbers mean when compared to the M1 Ultra Mac Studio: 2.4x faster project builds in Xcode 4.7x faster render performance in Redshift 9.8x faster time to first token performance in LM Studio 15.4x faster CopyCat ML training in Foundry Nuke Storage is fast, really fast Of course, with these machines built to work with and manipulate huge files, one roadblock to performance will be storage, right? Not on this Mac. Apple says it has deployed a new storage controller tech in the computers, which works with the speediest NAND memory it could find to deliver twice the read and write performance we got from the last generation of this system. This gives it plenty of horsepower for flinging files about, with Blackmagic’s Disk Speed test giving me exceptional results: 13,941MB/s write and 11,350MB/s read speeds. These speeds are indeed double the performance of the previous generation. It also means this Mac Studio can handle data transfers faster than almost anything out there, making it happy to handle multiple streams of uncompressed 8K RAW video, heavy compositing workloads and, of course, AI development, machine learning, or running your own on-premises AI models. Install the models you want to use and use them to your heart’s content. Run it headless if you like. I did. Apple Headless, no hanging I know a lot of you will end up wanting to run some kind of headless setup using this Mac. You’ll have it working furiously as your domestic or business AI server, chewing through your data, vibe coding opportunistic app creations, rendering video off your main Mac, and more. In my own little experiment, I found myself typing sentences for this review (this sentence, actually) on a Mac mini using a keyboard on a MacBook that happens to have the Mac Studio in its active window over vnc, and nothing ran slow. It means that if you run this Mac headless, it’s no slouch. Better yet, once you have the Mac set to run as a headless unit, you’ll be able to download LM Studio, install your choice of AI, and chat to your heart’s content. Your AI running privately and securely for you on your device, and — one more thing — it’s fast and responsive. What’s not to like? I used it to design and develop a capability test to put the Mac through its paces. Apple’s focus on AI is strategic, of course. Apple knows its hardware has pretty much occupied the AI development space, to the extent that almost any LLM you use was probably at least in part made on a Mac. AI is up there with CAD and medical imaging among the most demanding tasks you can do on any PC, let alone a Mac. And these Macs can handle all those tasks. I did want to try stringing four of these Macs together to run as an AI cluster, but at $12,000+ each that wasn’t going to happen. Buying advice Most of us don’t need this Mac. We probably never will — which is why cost is not the point here. This Mac is about performance, full stop, and that shows at every layer: a faster processor, blistering SSD storage, unified memory that scales to the task, and a storage controller fast enough to leave most other PCs gasping in the dust. From the software to the silicon, it feels like Apple’s engineers raided every high-end tech they had and crammed it inside this good-looking silver box. The real question isn’t whether this Mac is impressive — it obviously is — but if it’s impressive enough to justify an upgrade if you already own the previous model. Things get a little more nuanced if that is the case. The M5 Ultra is a genuine step up: the processor gains are significant, and the storage speed increase alone will matter to plenty of pro workflows. But last year’s Mac Studio was already so far ahead of most computers that “significantly better than the best thing available” doesn’t automatically mean “worth $12,000+ to replace.” My take: if you’re still on an older computer and need the extra processing power or storage throughput for demanding work, this machine delivers in spades. If you bought last year’s M3 Ultra model and it’s handling your workload fine, there’s no urgency. You’re not falling behind, you’re just not on the bleeding edge. All the same, I so wish AI-driven price inflation hadn’t pushed these systems quite so high in price. Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core , or follow me on BlueSky , LinkedIn , or Mastodon .
I’m old enough to remember when Macs were friendly little machines. While a little underpowered and equipped with some platform-unique foibles, they were really good at some things, highly secure, and had the best user interface of any PC. That was then; this is now. And while the platform is still unique, still highly secure, and still has the best user interface, Apple’s all-new Mac Studio is a beast of a machine. It’s incredibly powerful and would crackle with energy if it weren’t so energy efficient. TL;DR review No matter what pro workflow you follow, the M5 Ultra Mac Studio is more than you need. Longer review Apple hasn’t changed the design of the Mac Studio. It’s 7.7-inch square, 3.7-inch high aluminum box with two USB-C and one SDXC port on the front, and an array of ports — four Thunderbolt 5, two USB A, 10Gb Ethernet, HDMI, and headphone — at rear. It brings Wi-Fi 7 and Bluetooth 6, supports up to eight displays with up to 6K resolution at 60Hz, or four Studio Display XDR units. Basically, this one box can power a bank of displays, and for those using it in a live video production environment, Apple has also introduced generator locking (a.k.a. genlock) over USB-C, which lets the on-display image synchronize precisely with a pro video camera for the most precise editing and playback you’ll find. Pros and cons Pros Massive CPU/GPU power capable of handling the most demanding 3D and AI tasks. Neural accelerators and unified memory make it ideal for training and running large LLMs on-premises privately. Doubled read/write performance over the previous generation via a new storage controller and fast NAND. Inclusion of four ProRes accelerators (handling up to 33 streams of 8K) and genlock over USB-C for precise sync. High performance at low wattage; runs quiet and cool compared to high-end PC alternatives. Features Thunderbolt 5, Wi-Fi 7, and Bluetooth 6. Works exceptionally well as a headless AI or rendering server. Cons Extremely high retail price ($12,299 for the top spec). Neither the RAM nor the SSD can be upgraded or replaced by the user. No keyboard or mouse included, which feels odd for a five-figure machine. For my review, Apple loaned me a top-of-the-line M5 Ultra Mac Studio equipped with a 36-core CPU and 80-core GPU. It has 256GB memory, 4TB storage, and carries a retail price of $12,299. At that price and with those specifications, it is of little surprise that this is such a great computer. It’s also intended for some of the most demanding 3D and AI tasks you’ll find out there, rather than for people who spend their time writing about tech — though this machine can open multiple tabs in Chrome without stuttering, so unleashes real productivity boosts even at my level. Apple Let’s get one thing out the way That’s not to say some pro users lack reservations about it. Many are frustrated that neither the RAM nor SSD are user serviceable, while the lack of a keyboard and mouse in the box chimes oddly in a computer that costs five figures. In Apple’s defense, in these pro markets most users probably already own better keyboards and mice than you’d find in the box, while the massive performance benefits you’ll find with the M5 Ultra chip are in part realized because RAM in Macs is built to be part of the processor itself. That SOC design boosts efficiency, cuts heat dissipation, and makes for great performance at low wattage relative to other machines. Changing the way memory is installed on the chip would compromise Apple’s silicon architecture, which Apple doesn’t want to do. It’s all about the architecture The Mac is powered by a quad-die chip built using Apple’s UltraFusion process. Inside is an engineering marvel. The 80-core GPU brings neural accelerators, making these Macs truly phenomenal machines for building, developing, training, or running AI models. Apple’s unified memory tech is a godsend for tasks like these as it scales to handle… well, most anything you throw at it. Efficiently. Apple’s coupled chip speed with performance across the system. That means an advanced storage controller, superfast NAND, and really fast, next-gen SSDs for twice the read/write performance of the last model. This basically means the chip can grab big dollops of data and very swiftly push them around the system all the way from storage to display. Then there’s the rest of the advanced tech to consider, including the inclusion of four ProRes accelerators, which means my test Mac can deal with up to an incredible 33 streams of 8K ProRes video. This is a machine that can happily handle massive multi-camera editing, racing through dozens of high-res angles at once to field the perfect shot. Think live sports, concerts, and movie shoots. Not only this, but all those streams are at normal res, no conversion required — no sitting around waiting for low-quality proxy files. Apple A bicycle for several minds I’m thinking a full-flight video rendering data server in an 8-inch box that consumes perhaps 10 cents an hour at peak power (c. 480 watts). I’m also thinking of it as an on-premises AI system for me, the family, or any enterprise. Compared to the previous equivalent mode, the M3 Ultra, Apple says the chip brings up to 4.3x faster AI performance, up to 1.8x faster GPU performance, and up to 1.3x faster CPU performance. It’s also almost ten times faster than the M1 Ultra for AI. None of this is accidental. All of it is designed. This whole creation is architectural; it leans heavily into Apple’s software and hardware integration, which now also extends to the design of the processor itself. Making the RAM user-serviceable would limit the performance of the machine, which at this price and in this sector of the market seems a little counterintuitive. Sure, you can build yourself something pretty powerful using a Ryzen 9 chip that consumes 900 watts at peak and runs hot. Or you can put Apple’s silver box on your desk and barely hear a thing as it crunches through ‘god tier’ AI models your Ryzen can’t handle without additional GPU’s. All the same, if you want to configure your own memory you do have a choice — it’s just not a Mac. Choice is nice I know what I’d choose. Based on a weekend of using the Studio, I’ve found what it does is beyond most of the feeble tests mere mortals like me can cook up, so I thought you might want some benchmarks: Geekbench 7 Single-core CPU : 3,771 Multi-core CPU : 52,350 GPU (Metal): 360,019 GPU (OpenCL): 214,466 Cinebench CPU (Multi-thread): 18,052 GPU: 141,480 Putting these numbers into context, Apple explains what these numbers mean when compared to the M1 Ultra Mac Studio: 2.4x faster project builds in Xcode 4.7x faster render performance in Redshift 9.8x faster time to first token performance in LM Studio 15.4x faster CopyCat ML training in Foundry Nuke Storage is fast, really fast Of course, with these machines built to work with and manipulate huge files, one roadblock to performance will be storage, right? Not on this Mac. Apple says it has deployed a new storage controller tech in the computers, which works with the speediest NAND memory it could find to deliver twice the read and write performance we got from the last generation of this system. This gives it plenty of horsepower for flinging files about, with Blackmagic’s Disk Speed test giving me exceptional results: 13,941MB/s write and 11,350MB/s read speeds. These speeds are indeed double the performance of the previous generation. It also means this Mac Studio can handle data transfers faster than almost anything out there, making it happy to handle multiple streams of uncompressed 8K RAW video, heavy compositing workloads and, of course, AI development, machine learning, or running your own on-premises AI models. Install the models you want to use and use them to your heart’s content. Run it headless if you like. I did. Apple Headless, no hanging I know a lot of you will end up wanting to run some kind of headless setup using this Mac. You’ll have it working furiously as your domestic or business AI server, chewing through your data, vibe coding opportunistic app creations, rendering video off your main Mac, and more. In my own little experiment, I found myself typing sentences for this review (this sentence, actually) on a Mac mini using a keyboard on a MacBook that happens to have the Mac Studio in its active window over vnc, and nothing ran slow. It means that if you run this Mac headless, it’s no slouch. Better yet, once you have the Mac set to run as a headless unit, you’ll be able to download LM Studio, install your choice of AI, and chat to your heart’s content. Your AI running privately and securely for you on your device, and — one more thing — it’s fast and responsive. What’s not to like? I used it to design and develop a capability test to put the Mac through its paces. Apple’s focus on AI is strategic, of course. Apple knows its hardware has pretty much occupied the AI development space, to the extent that almost any LLM you use was probably at least in part made on a Mac. AI is up there with CAD and medical imaging among the most demanding tasks you can do on any PC, let alone a Mac. And these Macs can handle all those tasks. I did want to try stringing four of these Macs together to run as an AI cluster, but at $12,000+ each that wasn’t going to happen. Buying advice Most of us don’t need this Mac. We probably never will — which is why cost is not the point here. This Mac is about performance, full stop, and that shows at every layer: a faster processor, blistering SSD storage, unified memory that scales to the task, and a storage controller fast enough to leave most other PCs gasping in the dust. From the software to the silicon, it feels like Apple’s engineers raided every high-end tech they had and crammed it inside this good-looking silver box. The real question isn’t whether this Mac is impressive — it obviously is — but if it’s impressive enough to justify an upgrade if you already own the previous model. Things get a little more nuanced if that is the case. The M5 Ultra is a genuine step up: the processor gains are significant, and the storage speed increase alone will matter to plenty of pro workflows. But last year’s Mac Studio was already so far ahead of most computers that “significantly better than the best thing available” doesn’t automatically mean “worth $12,000+ to replace.” My take: if you’re still on an older computer and need the extra processing power or storage throughput for demanding work, this machine delivers in spades. If you bought last year’s M3 Ultra model and it’s handling your workload fine, there’s no urgency. You’re not falling behind, you’re just not on the bleeding edge. All the same, I so wish AI-driven price inflation hadn’t pushed these systems quite so high in price. Now please subscribe to my daily, human-curated Apple-related news headline feed at The Core , or follow me on BlueSky , LinkedIn , or Mastodon .