The Surface Laptop Ultra delivers on many of Microsoft's promises

Microsoft promises a device for developers, AI and gaming, and more than delivers.

The Surface Laptop Ultra delivers on most of Microsoft's promises

Microsoft promises a device for developers, AI and gaming, and more than delivers.

Over the past week, the laptop Microsoft sent me has been digesting books and movies, generating images from my imagination, and running a 70-billion-parameter AI model that can fit into most laptops. It also ran AAA games at a speed and fidelity that I thought was impossible just a year ago. Microsoft wants to sell you a computer for all your needs, and for once it doesn't seem like an entirely glamorous market. more

With the Surface Laptop Ultra, Microsoft has wrapped both arms around its new Nvidia RTX Spark platform, targeting the potential MacBook Pro buyer, or as the company calls it, the "high performance audience." The bet is that instead of a gaming desktop for fun and a developer laptop for work, buyers will pay more for a device that does both.

The Surface Laptop Ultra arrives at an odd time in personal computing. Recent industry figures show a 20 percent drop in total PC shipments in the third quarter of 2026. Between the RAMpocalypse that has blown up consumer prices and an industry focused on corporate money, there's not much new hardware to get PC enthusiasts excited.

Microsoft is not the only one who has introduced a laptop equipped with an RTX chip. Dell's XPS 16 Creator Edition and Asus' ProArt series (our own Ryan Whitwam's review will appear on the front page soon) are also betting on the chipset. The question when reviewing the Surface Laptop Ultra isn't whether it can achieve what it claims to be capable of, but whether there's finally a compelling reason for power users to switch to an ARM system, and whether it can still get through their day-to-day workloads without a problem compared to a more traditional x86 version or its biggest competitor in the segment, the MacBook Pro.

With a starting price of $2,599, it's not an economical alternative to Apple's silicon. This is advanced hardware and you will pay for it. Depending on the processor and memory options you choose, the MSRP can go as high as $5,899. Potential buyers should note that the lower-spec 24GB integrated memory model comes with Windows 11 Home, while the rest of the fleet comes with Windows 11 Pro. I got the 64GB model for testing, and I think that's probably the sweet spot for many users when it comes to balancing cost and performance.

I really have to question if 24GB is enough to get around when we know that Windows 11 itself is memory hungry. Does performance suffer when you have to dedicate a significant portion of that 24GB of integrated memory to VRAM for gaming? An audience question at the Surface Laptop Ultra review session was asked about the usefulness of the base model, and it didn't get much, other than a vague question. The answer was "locally useful for smaller models." Microsoft may have a more definitive answer to that question in a few months, when the software and drivers hit the hardware.

So have Microsoft and Nvidia teamed up to create the ultimate all-in-one laptop? Or is this industry once again placing the hardware shopping cart on the finished software horse?

The 15-inch Mini LED display might not be enough to entice people to go Ultra, but it's a compelling reason to add it to the list. The native 3270x2180 resolution looks simply stunning whether you're watching a 4K video, playing a game or staring at a wide desktop screen. I prefer the Advanced Color Profile, but the Vivid and sRGB modes also look fantastic.

The Surface Laptop Ultra's keyboard is exactly what you'd expect from a premium device like this. Great, if unremarkable. The keys have a shallow but satisfying press, the spacing is reasonable for my hands, and the whole package offers a very comfortable typing experience. I don't have many laptops to compare it to, but unsurprisingly, the closest is my MacBook Pro 2020. Microsoft knows what market they're going after here, and I wouldn't mind them taking inspiration from Apple in this department.

The touchpad is similarly capable. It's quite large, measuring 5.5 inches long and 3.5 inches tall, with excellent tracking accuracy during my testing time. Normally, one of the first things I do when I plug in a laptop for work is to plug in an external mouse, but over the course of several days of testing, I never touched the mouse drawer. Microsoft highlighted the trackpad's haptic feedback capabilities in native apps during the review briefing. I'm not sure if there are any updates left that I haven't received yet, but I can't remember a time when I noticed an extra mouse touch when I opened MS Office or anything. As it turns out, the trackpad isn't groundbreaking, but it's premium enough to be part of a great laptop.

Microsoft describes the included magnetic charging port and connector as "the world's first built-in USB-C magnetic charging experience." It's hard to begrudge the marketing department doing its job, but this claim struck me as more of a technical fact than an engineering feat. Apple's MagSafe charging debuted on the original MacBook Pro in 2006. The concept of a detachable magnetic charger isn't new, but technically Microsoft's claim is that this is the first USB-C interface on a laptop to match the magnetic port, not a dedicated connector.

Marketing claims aside, the magnetic connection works well. The port and charger naturally align like magnets, making plugging the laptop in to charge in a dark room effortless. Usually I'd hold the cord in one hand while trying to fumble around in it's dark for the phone in my pocket and remember the gesture of quickly turning on the flashlight. Now I just need to bring the cable into proximity and the magnets will do the rest. Perfect for when all I want to do is plug in the PC and go to bed.

Perhaps most important to magnetic charging is the cable's ability to detach when disconnected from a PC. Nothing ruins your day faster than watching a new laptop fly off the counter. With the cable disconnected, you have peace of mind that your laptop won't have to touch the ground if a startled cat or curious toddler comes into contact with the cable. I've never had this happen with one of my laptops, but that doesn't mean it doesn't happen. I simulated a few "accidental breaks" with low force and the cable cut dutifully as it should.

While I wouldn't say it's a must, it's certainly nice to have. And in this package it's certainly the kind of premium accessory you'd expect, but it's not going to move units on its own – not at these price levels.

According to Nvidia, the RTX Spark N1X is a Grace CPU combined with a Blackwell GPU. Unlike chipsets like AMD's Ryzen AI and Intel's Core Ultra series, the N1X lacks a dedicated NPU, instead relying on integrated memory and a heavy-duty Blackwell GPU to handle AI. Traditionally, Copilot+ certification hinges on a 40+ TOPS NPU, but Microsoft extends that label to the device's GPU-based AI performance anyway (more on that below).

The graphics benchmarks beep when they reach 40GB of usable RAM. These results certainly back up Microsoft's claims that the Surface Laptop Ultra can play modern AAA games. It just takes a different route to performance than we might be used to.

The artificial intelligence benchmark was a bit wrong. I don't have the ability to run an NPU-accelerated benchmark in Geekbench AI, as the software probably doesn't currently have a compatible backend NPU for this unique architecture. Oddly, Windows lists a generic NPU under Display Adapters in Device Manager, but Geekbench AI only provides the ONNX CPU test path. It is not clear if the public NPU listed by Windows is a virtual machine, some kind of driver exposure, or something else.

It's still not possible to get performance benchmarks on hardware that Geekbench can't properly see. It's still early days, but it's the kind of software/hardware incompatibility you hate when you want to see what your new computer can do. The CPU score on Geekbench AI isn't all that impressive, and bypassing Blackwell defeats the purpose of this hardware, so hopefully we'll get a better idea of ​​the full potential when the RTX Spark machines start working.

AI testing was currently limited to CPUs.

AI testing (if possible) with AI Dev Gallery

Microsoft recommends reviewers try out the new AI Dev Gallery app (preview) as a way to try out AI on the device. The software is currently in public preview, and users can pull their favorite models directly from Hugging Face for testing or workflow. The project's GitHub contains a helpful warning about models running on ARM64-based Copilot+ PCs, which Microsoft says they are. Or eventually will. Confusing, isn't it?

⚠️ Note: On ARM64-based Copilot+ PCs, make sure to build and run the solution as ARM64 (and not x64). This is needed, especially when running examples that are related to models like Phi Silica.

On paper, the AI ​​Dev Gallery should be a simple store for testing a bunch of different models against the new RTX Spark Surface Laptop Ultra launch. In practice, all of the Phi and Mistral conversational models in the program are CPU-only—no GPU or NPU paths are available, despite Blackwell sitting idle. GPU and NPU paths are available for some image generation options and embedded models, but it gives me the lingering feeling that we're using high-end hardware just to have something to offer the market. Oh, and it's worth noting that twice during my testing, I was browsing the AI ​​Dev Gallery model selection when the app suddenly stopped and then crashed back to the desktop entirely. There are still rough edges to smooth out.

Phi Silica, Microsoft's flagship LLM on-device and core part of the Copilot+ stack, is not accessible out of the gate. Accessing the API requires a developer to register with an associated identity program. These may seem like obvious steps to those already familiar with AI, but curious consumers may be disappointed when they arrive at a gateway advertised as a ready-to-run experience out of the box.

AI Dev Gallery Conversation Model (CPU Mistral 7B Instruct 0.2) - With the GPU off, this CPU-only model was still able to handle about 11 tokens per second. Certainly slower than the GPU route, but it's interesting that the CPU alone can handle smaller models at a reasonable conversation rate.

Image generation AI Dev Gallery (Stable Diffusion v1.4) — I tried randomly using the CPU at first. It didn't work so well, but once I switched to the NvTensorRT (GPU) option, I was able to generate screenshots from a single command in about three to four seconds with my stopwatch. The results aren't the best you'll ever see, but with additional queries and image matching I'm sure they could have improved. Whether you need local AI image generation in your daily driver is another matter.

I tried following Microsoft's instructions listed in the AI ​​Developers Gallery to run a CUDA-accelerated Phi model, but the process failed. After solving some command line problems, pip downloaded the correct Python 3.13 onnxruntime-genai-cuda package for Windows ARM64, but then failed because it didn't seem to have the onnxruntime-gpu dependency required by the distribution for this environment. Reports indicate that the CUDA installation path cannot be used in this setting.At this time, whether this is a limitation of the current package, a gap in the ARM64 package, or human error, I cannot be entirely sure. It was just frustrating that I felt limited in the space that Microsoft told me to play in.

I'm trying to get a picture of my dog.

The AI image has a quick rendering of a one-wheeled monster.

Frustrated with the current state of Microsoft's AI Dev Gallery app, I decided to try some experiments through Ollama running locally on the command line. I downloaded the Llama3.3-70b model. This is a 43 GB model - 10 times the size of the Mistral model we ran on the CPU in the AI ​​Dev Gallery, with 10 times the parameters (70 billion on 7).

If I were Microsoft, this is the kind of demo I'd want to show. This massive 70B model is unlikely to run on any individual laptop GPU, but Blackwell still manages 6.26 tokens per second. The fact that it was performed at all is remarkable, although I will say it was one of the few times Blackwell seemed to actually break a sweat. System fans were spinning 100% while Wayne's World recaps were being written. Considering that, even at full tilt, the fans don't make much noise other than high-speed air pressure.

When Microsoft talks about the high-performance audience, none of the modern high-end PC gamers are demanding. I'm old enough that 1080p/60Hz gaming still looks great, and given the current cost of video cards it's probably the best, but those who want zero input lag, 4K, ray tracing life are a special bunch. Whether or not the package offered with the Surface Laptop Ultra will be high enough for that audience is up to the gamer and their wallet, but this gamer is certainly impressed that these results are coming out of an ARM device.

Microsoft claims that 85% of its Game Pass library is currently playable on ARM, but I had trouble trying to find out if any titles had an original ARM port or ran on x64 through the Prism emulation lens. Windows Performance Fit gives you a glimpse of how the game is performing on your system, but not much else. In any case, if you are concerned with running games, this system should not be run too much. The full Nvidia DLSS feature set appears to be available, and ray tracing boxes can be clicked in the graphics options. Throw in an HDR-capable display, and the Surface Laptop Ultra covers most of the essentials for the modern gamer.

After the initial launch of Gears of War: E-Day saw the shader take 4-5 minutes to compile, I was concerned that this might happen every time, but subsequent loads were much faster. It's a game that features anti-cheat, the latest Unreal Engine, and every conceivable graphical bells and whistles you'd want from one of your company's prized franchises. Here it is, and it runs great on the Surface Laptop Ultra. While benchmarking, I noticed that the system reported 46.7GB of its 64GB integrated memory as VRAM available for gaming.

To its credit, Microsoft included a warning for consumers who might not be familiar with the Surface Laptop Ultra's new architecture: “Application compatibility: Surface Laptop Ultra runs Windows 11 on the Arm-based Nvidia RTX Spark platform. Many popular applications run natively on Arm. Programs not built for Arm can run through Prism, your own built-in remulator hardware drivers. Older plug-ins or tools, check app and driver compatibility before purchasing.”

Multiplayer titles with kernel-level anti-cheat measures may fail or show warnings under Prism emulation due to signature mismatches in system files. Every time I run Gears of War: E-Day, I get a warning message when I exit Windows. Easy Anti-Cheat seems to mark the .NET cache as "modified", though I can't tell if Prism's translation was directly responsible.

Additional benchmark results in Forza Horizon 6.

Beyond gaming, I wanted to see how video encoding works on the new hardware. I ran a real-world encoding test through HandBrake on a 2160p 60fps version of the famous Big Buck Bunny. The system processed a 10-minute video clip in 2:27 at 180 fps using x264 encoding via the CPU. This is the problem. NVINC or Nvidia HEVC encoding options are nowhere to be found, which means the Blackwell encoding engine is still not exposed. Software-wise, this isn't a problem with HandBrake, it's just a continuation of a familiar pattern. Hardware decoding via Media Foundation worked well, so especially the encoding aspect of things that hit a road.