Back at Computex, I wrote that RTX Spark could be Nvidia’s Apple Silicon moment: a beefy Arm chip in a slim, premium machine, aimed at pro users rather than at the everyday ultrabook crowd. Now that I’ve spent time with the platform, gone in-depth with experts at a recent event, and had the chance to test it for myself, it’s something that has to be seen to be believed.

The RTX Spark machines may well be expensive, starting at $2599 for a more ‘basic’ config, and going well into the thousands for a top-spec machine. But after living with the numbers, I don’t think that’s the interesting part.

The interesting part is that Nvidia has built a chip that does two jobs very well, straight out of the box. It games far better than an integrated GPU has any right to, and it runs serious AI models locally, which is where Nvidia really hopes you’ll be looking. On the GPU side in particular, this is a completely different game to most of the competition.

What RTX Spark actually is

In case you need a quick catch-up, here’s what Nvidia’s shiny new chip actually is. Nvidia calls RTX Spark a “superchip”, and the top N1X part is easily one of the beefiest SoCs I’ve seen in a laptop.

It pairs a 20-core Grace CPU, co-developed with MediaTek and split between ten Cortex-X925 performance cores and ten Cortex-A725 efficiency cores, with an integrated Blackwell RTX GPU. That GPU packs 6,144 CUDA cores, the same count as a desktop RTX 5070, so there’s plenty of raw grunt on offer. Power is flexible too, with a TDP that scales from 45W to 80W depending on the device, which looks mighty impressive on paper.

Alongside that sits up to 128GB of unified memory, running at speeds beyond 9,400MT/s, with up to 300GB/s of bandwidth through a 256-bit bus.

There are smaller configurations as well, with 24GB and 32GB of memory, plus a cut-down option with an 18-core CPU and 5,120 CUDA cores. The full-fat variant starts at 24GB and climbs to 128GB, with prices scaling accordingly.

Why those specs matter

A pool of memory this big isn’t only about guaranteeing the integrated graphics a decent slice of VRAM. It’s more about what’s arguably the whole point of these machines – AI. Local models are starting to close the gap on their cloud-based rivals for overall intelligence, which means you can now run a potent model on your own laptop without your data ever being sent to the cloud.

Take a very basic example. Claude Opus 4.7 Max, the model that brought Anthropic’s Cowork to the AI’s list of abilities when it launched in April, has an AA intelligence rating of 41. Qwen 3.8 Flash Next, which arrived in August and can run locally, scores 40. On intelligence alone, that shows local AI catching up with the data centres.

The other big argument is cost. An agent running on-device is effectively unmetered, so you don’t have to keep one eye on the bill.

Nvidia’s recent briefing used Claude Sonnet 5.5 Medium as its example. Running 61 million output tokens a month through it would cost $1200 in API fees, whereas pushing the same number through Qwen 3.8 Flash Next on an RTX Spark machine comes to around $20 in electricity. For intensive agent users, the pitch is that the hardware can pay for itself eventually.

Why it matters

Windows PCs have long split into two camps between classic x86 and the new kid on the block, Arm. Arm chips like Qualcomm’s Snapdragon X2 Elite offer efficiency and fast CPUs but more modest graphics (even if it’s getting better), while x86 machines deliver the graphical oomph and the software compatibility, but could fall short on efficiency. RTX Spark is an attempt to end that split by bringing Nvidia’s full stack to an Arm platform. That means ray tracing and DLSS for gamers, and CUDA for developers.

Rivals haven’t stood still – Apple has shown for years what a unified memory pool can do for outright grunt, and Qualcomm’s Snapdragon X2 Elite is the established Arm option on Windows. What nobody has offered, as far as I’m aware, is a Windows machine with this much shared memory and this much graphics muscle in a single chip, which has real consequences when it comes to performance.

Games are part of the story too. Arm has long been held back by anti-cheat systems that refused to run on it, but Easy Anti-Cheat and BattlEye now support Arm thanks to the work Nvidia has put in with Microsoft. Nvidia is promising 1440p gaming at more modest settings with RTX Spark, and while I haven’t tested exactly to its own recommendations, the results are still interesting.

RTX Spark isn’t necessarily a straight rival to either the current run of x86 or Arm laptop chips, certainly in my book. I maintain that Nvidia is chasing a higher class of user and performance than is offered by existing chips. For instance, Qualcomm’s Snapdragon X2 Elite and Elite Extreme SoCs thrive in the thin, light, all-day-battery machines most people actually buy.

By comparison, RTX Spark asks for a premium in return for beefy performance in a portable body, and the memory it needs won’t help the bill. That’s a different market, and Nvidia seems shrewd enough to know it. The software is lining up behind that audience too. Microsoft has made wholesale improvements to its OS to suit RTX Spark, meaning a range of games once blocked on Arm by anti-cheat now run, for instance, and Adobe has re-engineered Premiere Pro and Photoshop around the chip’s architecture to allow for much improved performance when paired with agentic AI workloads.

What it means for Windows

The bigger shift is happening in Windows itself. Microsoft is rebuilding the operating system around AI agents, and RTX Spark is the hardware it’s using to show what that looks like. As agents take over work that apps used to do, Windows has to decide how much they can touch, and Microsoft’s answer is a trust model as much as a smarter assistant.

Setting one up is meant to be easy. Microsoft is adding a dialogue box that opens on first boot that on-boards an AI agent into your machine, and you can keep track of updates via the Microsoft Store. Nvidia says the list includes GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy, with updates that streamline setup, tighten security and enable routing between local and cloud models. It’s a deliberate effort to move local agents beyond the command-line tinkering they’ve needed so far.

Containment is the other half. Microsoft Execution Containers, or MXC, fence each agent into its own sandbox, with Windows enforcing what each one can reach, and Microsoft says they’re now generally available on Windows 11. The approach rests on three ideas: containment, identity and manageability. Each agent carries its own credentials, so its actions are attributed to the agent rather than to the logged-in user, while the user, or an IT team using tools such as Intune and Agent 365, stays in charge as administrator. Support is broad. Microsoft lists agents including OpenAI’s Codex, GitHub Copilot, OpenClaw and LM Studio, and Anthropic is among those building on it too.

Then there’s what Microsoft calls hybrid intelligence, with models running on the device and reaching for the cloud only when needed. Windows ML provides the runtime across CPU, GPU and NPU, while a system called HydraFusion routes tasks between local and cloud models to stretch your token budget if required. In Copilot, a new toggle offers local file access with permission-gated actions, and taskbar search gains quick actions such as arranging windows or switching on do-not-disturb.

Memory makes the local possible, as long as you’ve got enough. Microsoft has pointed to a 284-billion-parameter DeepSeek V4 Flash model squeezed into 60GB. Nvidia’s own new Nemotron 3.5 Bolt model, built for agentic work on RTX Spark, packs 74 billion parameters into 24GB through 2-bit quantisation, with only nine billion active at any moment. A big shared pool is exactly what those models need, which is why both are pitched at RTX Spark.

The synthetic numbers

And now perhaps the most important thing – benchmark numbers. What’s here is a decent dive into how well it performs in a range of our traditional synthetic tests to help quantify the end result of years of Nvidia’s innovation.

In Geekbench 6.5, the 20-core chip scored 3,085 single-core and 19,313 multi-core. That’s a touch ahead of Intel’s Core Ultra X9 388H (3,003 and 17,285) and comfortably clear of AMD’s Ryzen AI Max+ 395 (2,808 and 13,271). Only Qualcomm’s Snapdragon X2 Elite Extreme is faster, at 3,711 and 21,940.

Cinebench R23 is an app built for x86, so it runs through emulation here and actually undersells RTX Spark a smidgen, especially in single-core, where it scored 1,323. Multi-core tells a better story, nonetheless, where 20 cores show their might.

The 16,762 result beats the X2 Elite Extreme (14,569) and the Ryzen AI Max+ 395 (12,785) and is level with the Core Ultra X9 388H (16,417). Native code shows what the chip can really do, though. In Cinebench 2026, it scored 555 and 6,710, which the benchmark’s own comparisons place roughly between Apple’s M3 and M4, and close to Intel’s best Panther Lake chips.

Graphics is where the chip stands apart. Its 3DMark Time Spy score of 11,763 is more than double the Snapdragon X2 Elite Extreme’s Adreno graphics (4,317) and well clear of Intel’s Arc B390 (7,189). It also beats Strix Halo, at 9,569. Those rival figures come from different power-limited designs, so treat them as a guide rather than a like-for-like. It isn’t a substitute for a machine with a high-end dedicated GPU, but for local AI, where memory and graphics matter most, it’s the result I care about most. It’s also what makes this a different game to the competition, whose integrated graphics sit a clear step behind in my tests.

Put together, the pattern is clear. The CPU is good rather than dominant, and Qualcomm’s single-core lead is a reminder that Nvidia’s processor isn’t the headline act. Rather, the GPU is, and the fact that it’s possible to get more than double the score Qualcomm’s iGPU garnered, and even higher than AMD’s Strix Halo, right out of the gate, is mighty impressive.

But can it game?

This is an Nvidia chip after all, and as much as AI is becoming the core focus, it would go amiss if I didn’t talk about gaming numbers. After all, Nvidia has packed the same number of CUDA cores as an RTX 5070, and the brand’s briefing made a lot about how it has implemented the full RTX stack with this chip, meaning DLSS, ray tracing and all the rest of it.

The honest answer about the RTX Spark’s gaming numbers is that it’s surprisingly good, with some limits. Next to a discrete RTX 5070 laptop GPU, raw gaming performance is a little constrained, but it’s comfortably playable. At 1080p Ultra, Cyberpunk 2077 averaged 69.73fps and Returnal 62fps. Step up to QHD, and they hold at 47.12fps and 46fps, which narrows the gap to that 5070 setup. Where it runs out of road is ray tracing. Cyberpunk at its Ultra ray-tracing preset managed 30.63fps at 1080p, and higher resolutions are tougher still.

That’s where Nvidia’s software earns its keep, though. As I say, RTX Spark gets the full RTX stack, including the second-generation DLSS Transformer model, which is a real step up from the original and lifted Cyberpunk to 73.25fps at 1080p in the rasterised test. There’s support for DLSS 5 too, with an update unlocking a model that’s twice as fast in supported games such as NBA 2K27, which launches this week.

Then there’s multi-frame generation, which uses AI to add up to three extra frames for every traditionally rendered one. Newer DLSS updates add a 6x multiplier, plus a Dynamic mode that scales the output to a frame rate target you set. With the maximum 4x multiplier I could access by default, Cyberpunk reached 218.12fps at 1080p, and 142.54fps with ray tracing switched on. One caveat is that frame generation needs a healthy base frame rate underneath it, otherwise, the picture can turn choppy and laggy. Think of it as a way to polish a decent experience rather than rescue a poor one.

It’s easy to get lost in the numbers, so step back for a second. Every one of those results comes from integrated graphics. A couple of years ago, that would have sounded absurd, and the fact that you aren’t giving up on efficiency in the same way you can with a proper gaming laptop is something to behold.

Since local AI is the pitch Nvidia and Microsoft are going for with RTX Spark, I also ran MLPerf Client, the MLCommons benchmark for on-device language models, on the GPU using both Microsoft’s Windows ML runtime and llama.cpp. With Llama 3.1 8B Instruct, Windows ML averaged 54.3 tokens per second with a 0.39-second wait for the first token, against 53.7 tokens per second and 0.32 seconds for llama.cpp. The smaller Phi 4 Mini Instruct is where things get quick, reaching 99.2 tokens per second on Windows ML and 86.3 on llama.cpp, with the first token arriving in under 0.2 seconds on either.

What struck me is that the type of task matters far more than which runtime you pick. Quick jobs like content generation barely register, with the first token appearing in about a tenth of a second or less. Summarising a longer piece is the real stress test, slowing Llama to 49.4 tokens per second on Windows ML and leaving you waiting around a second and a half for a reply to start. 

The SWE Agent scenario, an agentic workload, tells a similar story. Llama averaged an expected 58.6 seconds on Windows ML and 47.7 on llama.cpp, while Qwen 3 8B on llama.cpp landed at 51.0 seconds. For reference, Procyon’s Windows ML benchmark recorded a score of 578. Put it all together, and RTX Spark looks comfortable in the 8-billion-parameter class, with the better runtime depending on the job.

How local AI runs in practice

Benchmarks only go so far, so I put an agent to work. As much as Windows’ onboarding is meant to make getting started with agentic AI less daunting, my route was more hands-on. I installed the Hermes agent and chose a model that would fit in system RAM. For most of my testing, that was Qwen 3.8 27B, which sits happily in 32GB with the rest of the memory handed to the GPU. It can read images as well as text and is aimed at agentic work.

The task I gave it was to triage my inbox. Using Himalaya, I gave Hermes read-only access to Gmail through an app password and asked it to sift through pitches and requests and rank them. I asked it to give products in my remit with sensible deadlines a higher ranking, with scores saved to a local file.

It took about an hour. The result was a prioritised document, ordered by what it had identified and when each item was embargoed until, with the most important flagged and a next step suggested for each. It may be slower than a cloud assistant might be, although the prompt I gave it was very detailed and a model’s output is only as good as its input. Crucially, a local model did the reading, so none of my emails went to a cloud AI service, and it respected the read-only guardrail throughout.

In this instance, it feels like local agents are practical but patient. The hour isn’t a verdict on the chip, since the long, detailed prompt and the size of the model both played a part, but it does show that these workloads behave like batch jobs rather than instant answers, for now. For something like triage that runs while you get on with other work, that’s fine. For anything conversational, you’d want a smaller model or a faster setup. The upside is control: read-only access, a local file for the results, and a model that did its thinking on the device.

Does RTX Spark deliver?

Back at Computex, I called RTX Spark a statement of intent and the first step on the road to changing the Windows ultrabook game for good. Nvidia’s pitch was a big one: reinvent the PC, game like a proper gaming machine and run AI agents locally. Keynote promises don’t always necessarily survive a full benchmark suite, so it was interesting to see how much would hold up.

This one largely held up. The GPU is in a different league to its Arm rivals, games are playable well beyond what integrated graphics usually manage, and the local AI side is genuinely usable. That’s the combination Nvidia was selling at Computex, and it’s the one the numbers back up.

Not everything is settled, though. The CPU is good rather than dominant, and the biggest questions sit outside my test results – things like a laptop’s individual battery life and display specs, and most importantly, whether people will pay MacBook Pro money for a Windows machine, however impressive the silicon. With the first systems due very soon, we won’t have to wait long to find out.

So was it a first step? After the numbers, I think it was more than that.