实测一周 RTX Spark 笔记本：游戏、续航与本地 AI 表现如何

Nvidia’s RTX Spark N1X processor is finally here, and we’ve been collectively testing a small fortune’s worth of powerful laptops sporting this new “superchip.” RTX Spark is designed to be a high-performing Arm-based SoC for laptops and small desktop computers, rivaling the best from Apple, Intel, AMD, and Qualcomm. It has loads of cores, Nvidia RTX Blackwell graphics, and enough power and unified memory to run sizable local AI models.

The flagship placement is inside Microsoft’s Surface Laptop Ultra with a staggering 128GB of unified memory, which our own Tom Warren fully reviewed. Sean Hollister has been testing a 64GB version of the Asus ProArt P16, which I’ve also been testing in its top-spec 128GB config. Since I’m the most spoiled of the lot, I also have a 64GB Lenovo Yoga 9n 2-in-1 and a 32GB HP OmniBook Ultra in for testing.

All of these laptops have the same N1X chip with 20 CPU cores and 6,144 GPU CUDA cores, and they’re all wildly expensive — ranging from the HP OmniBook at $3,199.99 to the Asus ProArt’s $6,999.99. There’s an 18-core / 5,120 CUDA N1X chip in lower cost configs, but that only gets as cheap as $2,599.99 for a Surface Laptop Ultra or a smaller Asus ProArt P14 with 24GB of RAM.

These are big and expensive laptops built on new chips. They’re meant for developers, coders, and content creators who might also want the ability to use local AI and even play games using the latest Nvidia tech like DLSS 5. There’s a lot to test, and we have not had a ton of time to test it all. So below you won’t find an exhaustive scored review of any one laptop, but instead our impressions and thoughts from three writers using five laptops for a week.

Is Nvidia’s “superchip” really all that super? Are Apple, Intel, AMD, and Qualcomm officially on notice? Have any of us gotten any sleep?

Does it just work?

Antonio Di Benedetto: Windows on Arm is no longer a wholly subpar experience like it was a few years ago. Thanks in large part to the push around Qualcomm’s Snapdragon X chips, lots of development has been put into native Arm-based apps for Windows. And many x86 apps work fine via Microsoft’s Prism emulation layer.

Adobe has had native Arm versions of Photoshop and Premiere for some time now, and even though my precious Lightroom Classic is still an x86-based app I don’t notice any issues using it on Arm. At this point, you’ve gotta be relying on specialized software that’s exclusively x86 based for Windows on Arm to be a non-starter — or you want to freely play any games out there without any worry of incompatibilities.

Sean Hollister: Every so often, there’s that moment when an app takes longer to load than it should, and I wonder whether it’ll load at all. Even then, it typically loads. Occasionally, it takes more than one try. My Evernote notes disappeared a few times one day when I alt-tabbed away, but a force refresh fixed that. And apps don’t seem to respect my system preference for dark mode.

But I largely agree with Antonio. I made an RTX Spark my daily driver, with all my typical work and personal apps — Signal, Discord, Audacity, CapCut and Evernote, Irfanview, Slack, ScreenToGif and more — and nothing has failed me so far. I never once needed to think whether I was running an Arm or an x86 app. The real hitches came with gaming and AI, as you’ll read below.

Tom Warren: I’ve been using a Windows on Arm laptop with a Qualcomm chip daily for the past two years and things have improved to the point that I can’t tell if an app is native or not anymore. That’s the same story on the RTX Spark laptops, and I feel like Nvidia’s entry into this category will only improve the story for creative apps and gaming.

Can it game?

TW: Windows on Arm never looked like a serious gaming platform before now. While there still aren’t many native Arm games, Nvidia’s GPU helps bruteforce through the emulation layer. Your mileage in gaming is going to vary depending on the title, but for the most part I’ve found that medium settings, running at 1440p and DLSS performance, will get you to a stable 60fps in most games.

I mostly game on my desktop, where I’m spoiled by using an RTX 5090 on a 32-inch 4K screen every day with maxed-out game settings, but I didn’t mind the look of medium settings on a 15-inch display. I was able to run Black Myth: Wukong at 70fps on average at native resolution (3270 x 2180) with medium settings and DLSS performance mode enabled. I ran the same medium settings and DLSS performance in Cyberpunk 2077 and averaged 66fps, and Forza Horizon 6 managed an average of 57fps.

In some titles you may need to drop the resolution or settings lower, though. This definitely isn’t as good as a dedicated gaming laptop, but for something that’s roughly on the level of a MacBook Pro it’s more than good enough for gaming on the go.

SH: Like Tom says, the GPU power is there — but the CPU might hold back some of your games. I had plenty of power for a smooth 60fps in Returnal (at 1440p with high settings), Expedition 33 (at a DLSS “Performance” upscaled 1200p with high settings), Jedi Survivor (DLSS “Performance” upscaled 1200p with medium settings), and Shadow of the Tomb Raider (at native 1800p and highest spec).

But when I fired up Helldivers 2, even dipping down to a DLSS-upscaled 1080p was no guarantee of 60fps in the highest difficulty missions. There are just too many enemies on screen for the Spark to keep up. Star Wars Zero Company is even more CPU-bound — the Spark didn’t care whether I was maximizing or minimizing the graphics and screen resolution. No matter what I tried, I typically saw my framerate capped between 37fps and 39fps in battle.

Even ancient and low-powered games can have quirks. I tried the 2003 title Max Payne 2 and saw situations where framerate dipped to 50fps. I even saw a few hitches in the excellent UFO 50, a collection of 2D games, though a couple other 2D games ran great.

I’m also not sure if RTX Spark is suitable for VR yet. The excellent Titanfall 2 VR mod worked, and it seemed to run fine at first, but lost a little stability after a few minutes. That could be my Wi-Fi, because the driver for Valve’s Steam Frame wireless adapter wouldn’t install on the Spark. Neither did Steam’s streaming audio solution, so the audio came out of the Nvidia laptop instead of the headset.

Just so we’re clear, all of this is assuming you want a laptop maxing out its fans and positively dumping heat on your desk or lap. That’s not unusual for thin gaming notebooks, but it’s way louder and hotter than a good desktop or handheld.

	2560x1600	2560x1600	2560x1600	2560x1600
	Asus ProArt P16 RTX Spark (20-core, 6144 CUDA, 128GB RAM)	HP OmniBook Ultra RTX Spark (20-core, 6144 CUDA, 32GB RAM)	Framework Laptop 16 5070 (2025, 8GB VRAM)	MacBook Pro M5 Max
Black Myth: Wukong TSR	27	27	28	Not compatible
Black Myth: Wukong DLSS	41	44	45	Not compatible
Black Myth: Wukong DLSS FG	53	59	71	Not compatible
Cyberpunk 2077 Ultra	41	46	57	56
Cyberpunk 2077 Ultra RT & DLSS quality	30	35	44	42 (MetalFX Quality)
Cyberpunk 2077 Ultra RT & DLSS Frame Gen x2	53	56	68	75 (MetalFX Quality, FSR 3.1 FG)
Cyberpunk 2077 Ultra RT & DLSS Frame Gen x4	92	103	Not tested	Not compatible

AD: I’ve wanted a MacBook Pro-like laptop with full PC gaming chops for as long as I can remember, so I was hyped for these laptops to deliver the goods in graphics. The first night, I set up the Asus ProArt P16, with all 128GB of its ridiculously large pool of memory, and when I booted up Big Walk for my planned online evening with some friends my first thought was, “Well, that looks okay. I guess.” I know Big Walk isn’t some triple-A graphical titan, but it has some fairly demanding graphics for the “friend slop” genre. I expected the N1X to max it out, or at least get close, but not really.

I had to bump the P16’s 4K display down to 2560 x 1600, and instead of Big Walk’s Highest or Very high graphics presets, I brought it down to High to maintain a decently solid 60-ish fps. Does any of that really matter for this silly co-op game about solving puzzles with your buddies? Of course not. It looked fine, after all. But it was a sobering reminder that this “superchip” is no graphical magician.

As Tom and Sean point out, RTX Spark can do some decent gaming, but temper your expectations for medium or even low settings in some cases, and be ready to rely on DLSS and frame gen. There’s no denying that cheaper desktops and even laptops with discrete Nvidia graphics cards can easily beat out an N1X. Nvidia claims it’s like having an RTX 5070 Laptop GPU in an SoC, but in my testing it was somewhere between that and an RTX 5060 Laptop GPU — not a slouch, for sure, but if you’re a gamer first your money goes much further with traditional Nvidia graphics cards.

How’s the battery life?

SH: Nvidia’s partners saw fit to equip their first wave of RTX Spark laptops with some of the biggest batteries you can bring on an airplane with no questions asked — but even with a 99.9 watt-hour pack like the one in my Asus ProArt P16, these laptops won’t be winning awards for longevity. While Tom and Antonio spent their days testing every use case under the sun, I made the P16 my daily driver for normal work while watching the battery meter like a hawk. That mostly means writing stories in a web browser, and it’s not quite lasting the day.

My best run was 5 hours and 40 minutes of real-world work at 50 percent brightness, including an hour-long video call with Tom and Antonio to strategize our RTX Spark coverage plan. The next day, I got only about 5 hours from a full charge because I spent two of them editing this video in Capcut. The third time I drained the battery, it took 5 hours, too.

With web work, I’m typically seeing 13-17 watt battery drain on the laptop’s default settings, but that climbs to 20-30 watts with light video work, up to 50 watts during the (brief) Capcut video export, and I made it jump all the way to 120 watts playing Indiana Jones and the Great Circle. The battery couldn’t keep feeding at that rate — it slowly went down, and at 30 percent charge I saw a big performance hit as it dropped dramatically to 60 watts.

The good news is the Asus charges fast from a 140W USB-C PD power source. It took 20 minutes to go from dead to 50 percent battery, another 15 minutes to get to 70 percent, and a total of 1 hour, 15 minutes to reach 95 percent charge.

AD: I’ve had to run lots of benchmarks while plugged in, so I haven’t been able to spend as much time testing battery life as I’d normally like. Anecdotally, in lots of fits and starts, battery life seems solid. As Sean saw, you’re unlikely to get through a full eight hour work day unless you’re going out of your way to be efficient. And if you fire up anything heavy, like a local AI model (more on that in a minute), you won’t even make it to lunch. I’ll eventually do more thorough battery testing as I continue using these laptops, including our usual rundown tests for a bit of 1:1 comparisons.

Does local AI seem useful?

TW: Not yet. It all feels like a beta aimed at developers right now: from the install experience of AI agents flashing up command prompts and Visual Studio, to the fans blasting away for the most basic of queries.

I struggled to come up with many compelling use cases for the power of these chips for local AI work. A lot of tasks could simply be done far more quickly and without the laptop fans spinning up by using a free cloud model. The power here is really in leaving the agent running at full throttle to build something overnight, running tasks that developers would favor.

It’s also nice to have this all on a local device, without having to share emails and files with a cloud service and worry about the privacy aspects. Still, I would regularly close OpenClaw or Hermes and not let them run in the background all day, especially if I had given them permissions to access local files. I’m just not ready to trust AI fully, local or not.

Microsoft and Nvidia both envision a future where AI agents automate things on your PC, but I’m still not fully convinced. While there are some mundane tasks I could certainly automate, I like using my PC to edit photos or videos, write, and play games. I don’t want to sit there and watch a robot do all of that for me. That’s part of the joy of computing to me.

SH: I walked out of Nvidia’s Editor’s Day thinking I might finally have a reason to fire up local AI. I asked Hermes, running Qwen3.8-27B, to build me a daily news briefing much like the one in Nvidia’s on-stage demo. It took 10 minutes, and 10 percent of my entire 99.9 watt-hour battery, to finally produce a list of stories that took the cloud-based Claude (Opus 5.5) half as long — while utterly monopolizing my GPU and dramatically heating up the Spark laptop. The laptop stayed hot and a little sluggish as long as I kept Hermes open.

Claude also produced better results, with more detail and better sourcing. I guess I got the local AI tokens for “free,” I wouldn’t trust Claude with my email, and I probably could have used a smaller, lighter model. But I went with what the program recommended, and that was the result. I tried a clean reinstall of Hermes to make sure it wasn’t a misconfiguration, but this time, the prompt took even longer and used 60 percent of my battery. Ouch.

1/2

AD: I’m still early in my journey of learning to use local AI, and these RTX Spark laptops are obviously a viable — albeit, expensive — option. But even if you have 128GB of memory to use some chunky models, it’s still not a user-friendly experience for beginners. Microsoft’s Get Started intro in Windows 11 gives you easy one-click installs for Hermes Agent, OpenClaw, and Perplexity. (Which all have ESRB ratings for some reason?) But after that, you’re kind of on your own. I stuck with Hermes, since that’s what I was already using on the Mac Studio, but if you’re using it on a system with 32GB of RAM or less it just tells you there isn’t enough memory for its recommended local Qwen models — either routing you towards the cloud or fully relying on you to go fetch a smaller model on your own.

Using local AI has major privacy advantages over cloud models, but it’s definitely less convenient since local LLMs are much slower to process and more taxing on your system. Churning a local model through your laptop’s memory tears through any of these laptops’ large batteries, and the results from LLMs are still mixed. I burned through 30 percent of the Lenovo Yoga 9n’s battery in 36 minutes to get a Qwen 3.8 27B model to resize and add a white background to five product images in PNG format — and it cropped the images wrong and cut off the products. I got it to do it right after a couple follow-up prompts, and grabbing the power cable to prevent Hermes from killing the laptop in the process.

Is the hardware truly great?

1/8

Photo: Antonio G. Di Benedetto / The Verge

AD: These laptops are very expensive, but that also means the manufacturers are mostly bringing their A-games. Across the Asus, HP, Lenovo, and what I gathered from Tom’s thoughts on the Surface, the RTX Spark line is awash in great-to-excellent haptic trackpads, keyboards, screens, and speakers. The metal builds of the three I’m testing feel great to the touch. The OLED displays all look fantastic. And each of these are decently thin for the kind of workstation-class laptops that they are.

If I had to pick one for myself based on hardware alone I’d probably choose the Asus ProArt. It ticks the most boxes for me with its three USB-C ports, super bright display, superior inverted-T arrow cluster, and full SD card slot. The HP actually feels the nicest, with an impressively rock-solid chassis, but at nearly five pounds it’s the heaviest of the bunch, and its microSD card slot is much less useful to me.

As for the Lenovo, it’s actually been my sleeper favorite. It’s the only 2-in-1 in this crop I’ve been testing, and that just makes it more flexible (pun intended) and fun. You can use its large stylus on either the screen or the large trackpad, giving it lots more potential for graphics and design work. And tent mode with a large 16-inch 120Hz OLED is a delight for the occasional gaming session or watching shows and movies — it’s just not as bright as the Asus.

SH: I’ve only truly driven the Asus P16 so far, but I was impressed with the build quality of the whole spectrum of RTX Spark launch laptops at Nvidia’s Editor’s Day. The Asus P16 probably wouldn’t be my pick personally because of its extra glossy screen and oversensitive touchpad. I like to use my laptops in rooms with lots of ambient light, and find I have to crank up the brightness of this machine’s otherwise excellent calibrated OLED screen. Meanwhile, my palms keep tapping my mouse cursor on things I didn’t mean to tap, just because they accidentally brush the touchpad while typing.

The speakers and webcam also leave something to be desired, but the Windows Hello face login and proximity sensing are excellent. As soon as I walk away, the screen dims, and I’m logged right back in just by looking at my laptop. Despite the heat, Asus’ fans aren’t shrill — they’re very loud and noticeable while gaming, but at least it’s whoosh rather than whine.

TW: While I ran into some weird bugs with the Surface Laptop Ultra during testing, I can’t fault the hardware. The look and feel is very similar to a MacBook Pro, and the magnetic USB-C port really feels like something Apple should have done — it’s that good. The large haptic trackpad carries on the tradition of great Surface trackpads, with the added benefit of subtle haptic feedback when you move windows around or align objects in apps like PowerPoint.

I would have preferred an OLED display on the Laptop Ultra, but I found switching away from the default “Vivid” color option to “Enhanced” improved the contrast on the mini-LED panel. I also really appreciate the port section on the Laptop Ultra, especially that there’s a full SD card reader and even a USB-A port. Check out my full review for more thoughts.

Is it fast? Does it hang with Apple M-series chips?

	Asus ProArt P16 / Nvidia RTX Spark N1X / 128GB / 2TB	HP OmniBook Ultra / Nvidia RTX Spark N1X / 32GB / 1TB	Microsoft Surface Laptop Ultra / Nvidia RTX Spark N1X / 128GB / 1TB	MacBook Pro 16 / Apple M5 Max / 128GB / 4TB	Mac Mini (2026) / Apple M5 Pro / 48GB / 1TB / 10G Ethernet	Mac Mini (2026) / Apple M6 / 24GB / 512GB	MacBook Pro 14 / Apple M5 / 16GB / 1TB
CPU cores	20	20	20	18	18	12	10
GPU (cores)	Nvidia N1X (6,144 CUDA cores)	Nvidia N1X (6,144 CUDA cores)	Nvidia N1X (6,144 CUDA cores)	Apple M5 Max (40 cores)	Apple M5 Pro (20 cores)	12	Apple M5 (10 cores)
Geekbench 6 CPU Single	3170	3131	2864	4330	4294	4653	4208
Geekbench 6 CPU Multi	20343	20074	19030	29143	29104	21005	17948
Geekbench 6 GPU (OpenCL)	147750	146829	144352	145613	88531	61462	49059
Geekbench 7 CPU Single	2728	2694	2705	3693	3675	4034
Geekbench 7 CPU Multi	23659	23522	23445	35439	34021	22450
Geekbench 7 GPU (OpenCL)	117778	118629	115241	153190	88763	57560
Cinebench 2026 Single	560	565	534	734	715	809	736
Cinebench 2026 Multi	6638	6879	6759	8952	8059	5533	4486
Procyon image gen	1301	1480	1444	852	Not tested	Not tested	Not tested
Procyon computer vision 2	3650	3894	3893	Not tested	Not tested	Not tested	Not tested
Procyon text gen Phi 3.5	2457	2692	2555	4426	Not tested	Not tested	Not tested
Procyon text gen Mistral 7B	2184	2313	2233	3715	Not tested	Not tested	Not tested
Procyon text gen Llama 3.1 8B	2114	2263	2183	3498	Not tested	Not tested	Not tested
Procyon text gen Llama 2 13B	2023	2128	2078	3313	Not tested	Not tested	Not tested
PugetBench for Photoshop	Not tested	Not tested	13399	15716	17726	16739	12354
PugetBench for Premiere Pro (2.0.0+)	Not tested	Not tested	82107	154829	105041	75108	71122
PugetBench for DaVinci Resolve (2.0.0+)	Not tested	Not tested	90257	124942	85624		50882
Blender classroom test (seconds, lower is better)	15	15	15	15	24	33	44
Blender cosmos test (seconds, lower is better)	62	61	62	35	51	84	Not tested
Premiere 4K Export (lower is better)	Not tested	Not tested	2 minutes 18 seconds	1 minute, 10 seconds	1 minute, 51 seconds	2 minutes, 34 seconds	2 minutes, 47 seconds
Sustained SSD reads (MB/s)	7189.07	10373.83	14537.29	13638.91	12218.07	6759.86	7049.45
Sustained SSD writes (MB/s)	6719.82	9246.89	12582.13	17814.19	15420.95	7537.38	7317.6
Price as tested	$6,999.99	$3,199.99	$5,899.99	$8,149	$2,899	$1,299	$2,149

AD: Yes, the RTX Spark N1X is a powerful chip. These high-end laptop configurations we’ve been testing with lots of unified memory are meant to compete with the Pro and Max chips in Apple’s MacBook Pros. And while they can hang in some GPU performance tests, like tying the top spec M5 Max in a Blender render, Apple’s CPUs are much more powerful.

Apple has put lots of development into its in-house processors and the extra-fast “super cores” of its M6, M5 Pro, and M5 Max chips. And, frankly, nothing competes with those chips in single core performance — not even Nvidia’s “superchip.” (Seriously, can we stop with the word super?) But the part that stings most about the RTX Spark’s 20-core CPU is that even a lowly 12-core M6 Mac Mini competes with it in multicore performance — and a Mac Mini starts at just $900.

It’s awesome to see this kind of graphics performance without the need for a discrete GPU, allowing for thinner system designs with less noisy fans. I’m into it, just as I was with AMD’s Strix Halo APU — but like Strix Halo, Nvidia’s RTX Spark is a pricey part. And there’s no denying that Apple’s offerings are more well-rounded performers, so long as you don’t care as much about games.

TW: It’s definitely up there, but I think Qualcomm comes closer to Apple’s M-series chips on the CPU side. Qualcomm also had years of experimentation with Windows on Arm before it acquired Nuvia and managed to create a competitive chip with the Snapdragon X series.

I think Nvidia will need a similar amount of work to catch up on the CPU side for future RTX Spark chips. The development of the RTX Spark took at least three years, and I had heard from sources familiar with Nvidia’s plans that the chips were supposed to launch last year. So work has been ongoing, but the CPU side is still a little disappointing.

Nvidia picked MediaTek for its Arm-based CPU, a chipmaker that is better known for the Wi-Fi hardware in laptops, not the CPUs. I sure hope the next-generation is more powerful. The whole time I’ve been testing the Surface Laptop Ultra I couldn’t help but think the forbidden dream would be a combination of a Qualcomm CPU and Nvidia GPU. I can’t see that ever happening in reality, but it would certainly be interesting to see.

So what’s our take on RTX Spark so far?

TW: I think it will elevate interest in Windows on Arm for creative and gaming tasks, simply because we haven’t had Arm-based Windows laptops that are really capable of running games until now. Nvidia is targeting creators and developers with its first round of RTX Spark laptops, but there’s definitely a gap here for a more affordable version focused on gaming. Perhaps that’s the N1 chips we’ve heard rumors about.

SH: Nice, but too pricey. Too hot and loud when you’re doing the things that would theoretically justify that price. I care what they stand for, and the direction they’re pointed, far more than I care to own one.

AD: I love that there is now even more competition in the laptop space, but the super high prices makes RTX Spark laptops mostly unattainable. I want to see lower cost options come around, as well as an eventual second generation with higher CPU performance to hopefully close the gap further with Apple. I also hope this helps push development for Arm-based and Arm-compatible games forward. Maybe it could even one day mean easier porting of games to macOS. Wouldn’t that be a fun outcome?

As cool as it will be to keep testing these super high-end RTX Spark models, I’m most hyped to get my hands on a base Surface Laptop Ultra or Asus ProArt P14 with 24GB of RAM and see how they stack up to the entry-level 14-inch M5 MacBook Pro. That’s my personal sweet spot for Apple’s lineup — the people’s MacBook Pro. (Well, okay, less these days now that it starts at two-grand.)

It’s an exciting time in computing when it comes to chip choice, but it’s also an absolutely cursed time thanks to RAMageddon and sky high prices. I’m constantly wondering what laptops like these RTX Sparks or even Googlebooks might have cost if AI demands didn’t drive up the prices of RAM and storage. I have no idea who is going to spend up to $7,000 on these laptops when PC sales as a whole seem on a downward trend, but it’s cool they exist.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

• Antonio G. Di Benedetto
• Sean Hollister
• Tom Warren