I won't lie: I stuck my nose up at the Mac Studio M3 Ultra. I knew it would be powerful, but with a configuration pushing into a fifth digit on pricing, I struggled to think of a situation where the Mac Studio would make sense. So, when I tell you that the Mac Studio M3 Ultra is fantastic, know that I mean it.

It's a machine that can turn a skeptic into a fan, an unbeliever into an evangelist. No, it's not for everyone, and the $10,000+ configuration options are only for a select few. But the Mac Studio manages feats that simply aren't possible elsewhere, all while showcasing world-class design, performance, and power efficiency that you just don't see elsewhere.

About this review: Apple provided the Apple Mac Studio M3 Ultra for this review. The company had no input into its contents.

Mac Studio M3 Ultra pricing and availability

Apple offers up two flavors of the Mac Studio right now. You can either get it with an M4 Max, which has configurations starting at $2,000, or the M3 Ultra, which starts at $4,000. Apple's naming scheme at the high-end is a bit confusing, but the Ultra chip is essentially two Max chips fused together. So, the M3 Ultra is based on an older architecture compared to the M4 Max, but it's a much more capable chip overall.

The standard configuration comes with a 28-core CPU, 60-core GPU, and 32-core Neural Engine, and alongside the M3 Ultra, you get 96GB of Unified Memory and 1TB of storage for the $4,000 base configuration. With this version of the M3 Ultra, you can pack in 256GB of Unified Memory for a $1,600 upcharge, along with up to 16TB of storage. Apple charges a premium as you go higher in storage capacity. For instance, 4TB of storage is an extra $1,000, but 8TB of storage is an extra $2,200.

That's one of the major downsides of the Mac Studio. Storage is usually relatively inexpensive and easy to add to a PC, but you have no options for upgrading the internal storage of the Mac Studio beyond spending potentially thousands when you check out. I'd recommend getting a handful of external SSDs ready.

There's a second configuration of the M3 Ultra available, however. For an extra $1,500, you can get 32 CPU cores, 80 GPU cores, and a 32-core Neural Engine. This chip also unlocks up to 512GB of Unified Memory, which is $4,000 more expensive than the base configuration with 96GB. The storage capacity and pricing doesn't change with this version of the M3 Ultra.

The unit I tested was just short of maxed-out, packing the 32-core M3 Ultra, 512GB of Unified Memory, and 8TB of storage. This configuration will run you $11,699 at the time of writing -- don't worry, I've pulled out the white gloves every time I need to touch the thing. Pushed to the limit with 16TB of storage, the Mac Studio can cost as much as $14,099. That's a ton of money, no doubt, but it needs some context.

The Mac Pro, which tops out with the 24-core M2 Ultra, 192GB of Unified Memory, and 8TB of storage, will run you $11,799 -- plus an extra $400 if you want wheels. Even building your own desktop that can compete with the M3 Ultra Mac Studio will run north of $10,000 if you factor in a proper 32-core chip like the $2,000 Threadripper 7970X and a workstation GPU like the $5,400 RTX A6000. Yes, the Mac Studio with an M3 Ultra is very expensive, but the price really isn't unreasonable for what you're getting -- short of storage, where Apple has put some arbitrarily high prices for extra capacity.

Simple but potent design

But who's really surprised?

It's easy to forget how much power is packed inside the Mac Studio. It's a chunky silver box, significantly bigger than the Mac Mini design it's based on, but it clocks in at just 8 pounds. This is a PC that can make even a tricked-out desktop with server-grade components green with envy, but it's not some 60-pound behemoth that you'll struggle to pick up. I don't like that you can't upgrade the Mac Studio, especially when it comes to storage, but Apple has a very compelling argument for locking off user access given how petite the Mac Studio is in relation to how much power it offers.

Unboxing the Mac Studio is unsurprisingly seamless, with Apple carefully crafting the packaging to make the setup as frictionless as possible. The Mac Studio comes out, the soft-braided power cord is underneath, and beyond that, you're off to the races. I was able to set up the Mac Studio, Apple's Studio Display, and a Magic Keyboard and Trackpad in no more than 10 minutes.

Up front, you get a pair of Thunderbolt 5 ports along with an SDXC slot rated for up to UHS-II SD cards. Around the back, you get an additional four Thunderbolt 5 connections, 10-Gigabit Ethernet, two USB-A ports rated for up to 5Gb/s, an HDMI 2.1 output, and a 3.5mm headphone jack. The Thunderbolt 5 connections are what really matter here. With them, you can essentially add anything you want to the Mac Studio, be it monitors through DisplayPort 2.1, additional storage, or high-bandwidth external accessories like an audio interface.

Unfortunately, you don't get quite as cutting-edge connectivity wirelessly. You have access to Bluetooth 5.3 and Wi-Fi 6E. Bluetooth 5.4 is available, though it doesn't actually offer anything for you that isn't already available in Bluetooth 5.3. It's the Wi-Fi standard that's behind. Wi-Fi 6E is super fast today, but I would've liked to see Wi-Fi 7 support given that everything else in the Mac Studio is on the bleeding edge of what's currently available to consumers. It's not relevant now, but it could be a few years down the line.

By the numbers

Cute flagship you got there

Here's what you need to know about the performance of the M3 Ultra. Short of something like a Threadripper, you won't find any PC that offers as much power as the Mac Studio does with the M3 Ultra. I have some benchmarks here to compare the M3 Ultra to flagship CPUs like the Ryzen 9 9950X3D and Core Ultra 9 285K, but there isn't much of a comparison at all. The M3 Ultra, unsurprisingly, is in a completely different class of performance.

The numbers you can see above speak for themselves on that front. In multi-core CPU tests like Cinebench and Geekbench, the M3 Ultra steals the show, which shouldn't come as a surprise. The chip offers, at minimum, a 28-core CPU, and the unit I'm testing here has 32 cores. These desktop chips from AMD and Intel never stood a chance. A more apt comparison is to something like the $2,000 Threadripper 7970X, which also comes with 32 cores. And yes, you will see higher multi-core performance with AMD's 32-core monster. Apple still comes out way ahead when it comes to single core speeds, though, and that balance is what makes the M3 Ultra so impressive.

The CPU is only half of what's going on inside the M3 Ultra. You also get up to an 80-core GPU. Based on Geekbench, the GPU is powerful, though it's not offering anywhere close to the raw CPU power on display. In OpenCL, you're looking at performance somewhere around an RX 7800 XT or Titan RTX. With Metal, there's nothing faster than the M3 Ultra. It wins by default in that category, however, considering the latest non-Apple silicon to show up with Metal support comes from AMD's RDNA 2 range.

Putting the performance into a bit of context, the numbers for Premiere Pro and Photoshop look solid. The M3 Ultra isn't shooting super far ahead of the x86 competition, and in some cases, it's even taking a backseat. That's largely due to the GPU, though. The numbers here were gathered with a desktop RTX 4080 for the AMD and Intel chips. It's important to remember that, at the end of the day, the M3 Ultra essentially has an integrated GPU. The fact that it's even in the same conversation is mighty impressive.

These numbers are all well and good, but they don't really showcase how powerful the M3 Ultra really is. I wanted to simulate a power user scenario. I threw the normal benchmark process out the window and intentionally loaded up several apps to see how the performance would hold up when you're multi-tasking. And that's where the M3 Ultra really shows some power.

You can see the numbers above. There's a stark drop compared to the numbers I showed higher up, but the chip still offers very similar (and sometimes better) multi-core performance compared to flagship x86 CPUs. But here's why these numbers are so impressive. While running these tests, I had the following applications open:

• 7 Docker containers
• Multipass running two Linux VMs
• Parallels running a Windows 11 VM
• Steam downloading Baulder's Gate 3

None of my VMs were suspended or running in the background. That's right, the M3 Ultra delivers performance that rivals flagship x86 CPUs even when it's loaded up with multiple Docker containers, VMs, and downloads. That's kind of insane, and it's a perfect demonstration of what kind of performance you can expect if you're able to leverage all the horsepower in the M3 Ultra.

Surprising power

Powerful, but not power-hungry

The Mac Studio with the M3 Ultra is powerful, which is shocking considering how little power it uses. Using a custom script, we were able to track the power usage of the CPU, GPU, and ANE numbers reported by the Mac Studio over time, and compared to what's happening in the world of x86 chips, the results are extremely impressive.

That's immediately clear when you look at the CPU stress test above. Again, using a custom script, we maxed-out the CPU utilization with a full workload. The M3 Ultra climbed up to 90 watts and remained there. That's crazy. Chips like the Ryzen 9 9950X and Core Ultra 9 285K can spike up to 200W in some cases, while a proper 32-core Threadripper 7970X can easily climb into the 300W range during peak power usage. The power metrics aren't as granular as what you can find with third-party Windows apps, but the M3 Ultra is orders of magnitude more efficient.

The M3 Ultra doesn't remain at that power draw for long, either. In a more practical workload, you can see that the M3 Ultra mostly stayed under 60W during the few minutes it took to compile the Ladybird browser. There are few spikes up to 80W, but this is a very CPU-heavy workload, and the M3 Ultra is able to manage excellent performance while barely sipping power.

That's CPU power, but what about the GPU? Running DeepSeek R1 locally, which took up over 400GB in memory, the Mac Studio never topped 70W. This is an extremely demanding workload with a massive LLM running locally, and the Mac Studio doesn't even climb into triple digits. It's absurd in the best way possible.

Even with so much power on tap, the Mac Studio never made a peep during my testing. There was a bit of hot air exhausting out the back after running a 400GB+ model of DeepSeek -- more on that in the next section -- but I never so much as heard a fan spin. Part of that is the excellent power efficiency of the M3 Ultra, but it's also Apple's internal design, which is really working wonders here.

Redefining an AI computer

Copilot who?

One of the more interesting aspects of the Mac Studio is its usefulness for AI. Sure, any powerful computer is capable of running AI workloads, but the M3 Ultra puts a different spin on what an AI computer can look like. That comes down to Apple's Unified Memory. The Mac Studio with an M3 Ultra supports up to 512GB of Unified Memory. You can pack that much memory into something like a Threadripper rig — consumer flagships top out lower — but it's that "unified" bit that's special. Its 512GB pool is available to both the CPU and GPU, meaning that your RAM doubles up as VRAM if you want it to. That's something you can't get on a DIY desktop unless you're willing to shell the price of multiple tricked-out Mac Studios for some enterprise-grade GPUs from Nvidia.

Of course, we've seen this approach across a multitude of devices. Gaming handhelds often offer the ability to increase the VRAM at the cost of actual system RAM, and it's a somewhat similar approach. However, the difference is that all of these components are on the same board. "Unified" memory is truly unified, where the RAM works just as well for the GPU as it does for any other case. After a quick Terminal command (you don't need a command if you've updated to macOS 15.4) I was able to give the GPU access to 464GB of memory. The command we used was:

sudo sysctl iogpu.wired_limit_mb=475136

The reason this is necessary is that GPU memory is everything when it comes to AI models. While you certainly can try running a model like DeepSeek R1 from your PC's storage, the constant reference back to the included storage for processing inputs means that you'll struggle to get any kind of response in a timely fashion, and having that much memory on tap for the GPU is practically unheard of outside the data center. Even the fastest SSDs pale when compared to the speed of RAM, with the fastest PCIe Gen 5 SSDs topping out at around 11000MB/s in read speeds. In contrast, DDR5 at 8000MHz can transfer at a rate of 64GB/s, or nearly six times faster.

As a result, running a model as large as DeepSeek R1 is genuinely possible, and once it loads the entire model into RAM, its performance is nothing short of impressive. This is the kind of power that you might expect requires a high-end server PC (and, in fact, the hardware required to run this model is baffling), yet here it can run on a machine smaller than most consumer-grade PCs.

You don't have to just take my word for it, either. Above, you can see DeepSeek R1 running locally on the M3 Ultra Mac Studio. Sure, it's the Q4_K_M 4-bit quantized version of DeepSeek R1, but there's very little quality lost compared to the full-sized reasoning model without any quantization. Here's a quick primer:

• Qx: This refers to the quantization level. It's how much memory is used to store the model's weight.
• K: This refers to k-quantization, meaning that different bit widths for value are used in order to optimize memory usage.
• M: This refers to the block size used for quantization and can be S, M, or L, meaning small, medium, or large.

We are using the 3rd best DeepSeek R1 model here, with the only two above it being Q6_K and Q8_0. Those other models are too large to fit in the memory of the Mac Studio (coming in at 550.80GB and 713.29GB, respectively), but the performance loss is minimal. DeepSeek R1 is an already advanced model, so the quality lost still means that it's an incredibly competitive model when compared to the likes of ChatGPT. And I can't stress this enough: it's running locally​​​​​​. You could enable LM Studio's server mode and access this model via an API like you would ChatGPT, except you wouldn't even need to pay for it because it's yours.

When we think of the phrase "AI computer", typically we think of things like Copilot-powered PCs and laptops. This completely recontextualizes that conception, but it's something that just isn't possible with any mini PC or laptop aside from the Mac Studio. To even get close to this level of performance, you'll be spending $8,000 a pop on dozens of 48GB RTX 6000 Ada GPUs. This isn't a practical use case for everyone, and you're looking at a $5,500 upcharge for 512GB of Unified Memory compared to the base M3 Ultra Mac Studio. But for those who need it, there's nothing else that comes close.

There are times when this much RAM is required in a development environment. Developers who require multiple containers or virtual machines running can often require a significant amount of RAM, which is why these machines are offered with so much in the first place. I'm not saying you should go out and buy this $12,000 machine in order to run a local AI model, but what I am saying is that in the cases where you need one of these machines anyway, this is an incredible additional perk that you wouldn't be able to get on any other machine. Your high-end development workstation with 512GB of RAM still couldn't run this model locally; it's only thanks to Apple's Unified Memory that you can here.

I'm a believer

The Mac Studio isn't impressive because it's fast. You drop five figures on any computer, and it better be fast. It's not impressive because of the hardware you can cram inside. And it's not impressive because of Apple's user-friendly aura. It's impressive because it can pull off feats that other computers simply can't, and especially not at this price with the convenience of a prebuilt PC. Unified Memory makes the impossible possible for AI workloads, and Apple's careful power management allows you to get flagship CPU performance even when slamming the M3 Ultra with other applications.

But you need a reason to pick up this much power in this small of a package. The expanded Unified Memory helps a lot, but the M3 Ultra is wasted money if you aren't able to fully leverage its capabilities. This is a machine for developers, data scientists, and extreme power users. If you don't fall in that camp, you're spending too much. For those who do, though, there's nothing better than the Mac Studio.