Is an NVIDIA RTX Spark laptop worth it? Who it is for, and who should buy a MacBook

NVIDIA has been promising a Windows answer to Apple’s M-series chips since the RTX Spark reveal at Computex. The first laptops are now on sale from ASUS, Dell, HP, Lenovo, Microsoft and MSI, and the first serious reviews are in. We are about to buy one of these ourselves, for one reason: running local AI models on a Windows laptop. So we read and watched every test we trust before spending the money, and this guide is the answer we came to.

What NVIDIA RTX Spark actually is

RTX Spark is not a graphics card in a laptop. It is a whole computer on one chip. The N1X superchip, developed with MediaTek, combines three things that normally sit apart:

• An Arm CPU. NVIDIA calls it Grace, with up to 20 cores. Hardware Canucks identified the cores as Arm Cortex-X925 and A725 designs.
• A Blackwell RTX GPU with up to 6,144 CUDA cores, the same generation as the RTX 50-series.
• Up to 128 GB of LPDDR5X memory that the CPU and the GPU share.

NVIDIA rates the chip at 45 to 80 watts in laptops and quotes up to one petaflop of FP4 AI performance. The laptops run Windows 11 on Arm, the same version of Windows that Qualcomm’s Snapdragon laptops use.

One confusing detail: NVIDIA sells two different chips under the same N1X name. The smaller version has 18 CPU cores and 5,120 GPU cores and ships in the entry configurations, usually with 24 or 32 GB. The bigger version has 20 CPU cores and 6,144 GPU cores and is the only one you can get with 64 or 128 GB. When a laptop is advertised “from $2,599”, it is the small chip. For local AI, it is the big one you want.

Why unified memory is the whole point

The reason RTX Spark matters is memory, not speed. c’t 3003 explains it best. A strong local language model such as the 27-billion-parameter Qwen model they tested needs about 14 GB even when it is heavily compressed. Add enough context for a coding agent to remember what it is doing, and you are past 18 GB for the model alone.

A normal laptop GPU has 8 GB of fast video memory. Anything that does not fit has to spill into the regular system memory, which is far too slow for the GPU to work from. The model either does not load or crawls.

On RTX Spark, the GPU can use most of the shared pool. With 64 GB, a 27B model plus a long context sits entirely in fast memory the GPU can reach. Apple has worked this way since the M1, and AMD does it with its Strix Halo chips. What is new is that it now exists on Windows with NVIDIA’s GPU and CUDA software stack, which most AI tools are built for first.

It is also worth knowing what RTX Spark is not. It has nothing to do with the small NPU in Copilot+ PCs. The language models run on the GPU cores, which is where the speed comes from.

How fast local AI really is

The independent numbers are good, but not magic. c’t 3003 tested the ASUS ProArt P14 with llama-bench and no extra tricks:

• Qwen 27B (dense model, compressed): about 14.5 tokens per second. That is a little faster than Apple’s base M4 and M5 chips and AMD’s Strix Halo, and a little slower than an M5 Pro.
• Qwen 35B-A3B (mixture-of-experts model): almost 61 tokens per second, because only part of the model is active at a time.
• Prompt processing: when a coding agent reads 50,000 tokens of context in one go, raw GPU power counts, not memory bandwidth. Here RTX Spark clearly beat both Strix Halo and the M5 Pro, at only 40 to 60 watts.

Hardware Canucks compared the Surface Laptop Ultra with 128 GB against Apple’s MacBook Pro line. On smaller models it was 9 to 16 percent behind the M5 Pro. On the 27B model it was 10 percent ahead. On a 70B model it fell about 8 percent behind again, while the M5 Max, with twice the memory bandwidth, was almost twice as fast. Against an HP ZBook Ultra with AMD’s Strix Halo and the same 128 GB, RTX Spark was 60 percent faster on the 70B model.

The honest summary: for local language models, RTX Spark plays in the same league as an M5 Pro MacBook Pro, and it is the fastest option on Windows today. Apple has since put the M6 MacBook Pro on pre-order, so the Mac side of these comparisons will move again.

Speed is also not the same as quality. In c’t 3003’s coding-agent test, the slower 27B model built a clean, correct website, while the fast 35B mixture-of-experts model added made-up content. The number of tokens per second is only half the story.

Where CUDA makes the difference

On a Mac, language models run very well through tools such as Ollama, LM Studio and llama.cpp. So for chatting with a local model, CUDA is not the deciding factor. It matters for everything that was built for NVIDIA first: many image and video AI tools, research code, and developer workflows that assume an NVIDIA GPU. c’t 3003 names ComfyUI as the clearest example.

Linus Tech Tips showed the other side of the pitch: local AI agents. NVIDIA and Microsoft are building sandboxing and per-agent permissions into Windows, so an agent can work on your PC without full control over everything. Linus gave the local Hermes agent a multi-step job, building a system-monitoring dashboard and installing NVIDIA’s apps, and it finished on its own after a little over an hour. That is the use case NVIDIA is really selling.

Where RTX Spark falls short

This is the part the launch slides leave out, and the reviewers were clear about it.

• CPU performance. The Arm cores are an older design. In Cinebench, Hardware Canucks measured RTX Spark about 17 percent behind the M5 Pro in multi-core, and GeekaWhat’s single-core result was 464 against 740 for the M5 Pro.
• Adobe apps. This is the biggest weak spot. In Hardware Canucks’ tests, a Photoshop beta was 40 percent slower than on the M5 Pro, Lightroom Classic still runs through emulation and took more than three times as long to export RAW files, and the Arm version of Premiere Pro still lacks hardware export and several RAW video formats. DaVinci Resolve, on the other hand, runs well.
• Battery life. Hardware Canucks found the Surface lasting 20 percent less than a 14-inch M5 Pro in web browsing, and losing 26 to 40 percent of its performance on battery. Notebookcheck saw stable performance on battery with the ASUS ProArt P16, so this varies by model, but none of these laptops matches a MacBook away from the plug.
• Gaming compatibility. Many games run well, and Cyberpunk 2077 reached 63 fps at 1080p Ultra in c’t 3003’s test. GeekaWhat found the GPU close to a laptop RTX 5070 at 1440p. But x86 games run through Microsoft’s Prism emulation, and Hardware Canucks had 3 of 10 games refuse to start, including one blocked by anti-cheat.
• Fan noise and first-generation rough edges. c’t 3003’s ASUS ProArt P14 spun its fan up on an idle desktop right after boot. Several reviewers also noted that drivers and app support are still catching up.

None of these are reasons to dismiss the platform. They are reasons to be clear about what you are buying it for. This is a first generation, and software support on Windows for Arm has been improving steadily. But you should buy for what works today, not for what an update might bring.

Which configuration to buy

The configuration matters more than the brand.

• Get the 20-core chip. The 18-core version is limited in memory and GPU power, and the agent workflows NVIDIA advertises need the big chip. Linus Tech Tips made the same point.
• Get 64 GB. c’t 3003 hit the limit with 32 GB as soon as a model ran next to a browser and a creative app, and recommends at least 48 GB. Memory is soldered on every one of these laptops, so you cannot add more later.
• 128 GB only if you need it. It lets you load 70B models comfortably, but it adds thousands of dollars. For most people experimenting with local AI, 64 GB is the sweet spot.
• Check Windows on Arm compatibility first. Your VPN client, security software, printer drivers, dongles and any in-house tools should be tested before you commit, especially on a work laptop.

Prices for the 20-core, 64 GB versions currently land between roughly $4,500 and $4,800 in the US, depending on the brand. We compare the three we shortlisted, the ASUS ProArt P16, the Microsoft Surface Laptop Ultra and the Dell XPS 16 Creator Edition, in a separate head-to-head.

Who should buy an RTX Spark laptop

Buy one if you want to run serious local AI models on Windows, use tools that expect CUDA, build or test AI agents, or want one portable machine that does local AI, development and some gaming. In that role it has no real competition on Windows right now.

Buy a MacBook Pro instead if you spend most of your day in Photoshop, Lightroom or Premiere, if battery life is a priority, or if you do not need Windows. At the same memory size, the MacBook Pro runs local language models at a similar speed and does everything around them better.

Buy an x86 laptop instead if gaming or maximum compatibility comes first. An Intel or AMD laptop with an RTX 50-series GPU runs every game and every Windows app natively, for less money. AMD’s Strix Halo laptops, such as HP’s ZBook Ultra, also offer up to 128 GB of shared memory on x86, just with less AI speed.

Use the computer you have if your local AI plans are mostly experiments. A Mac mini with enough memory runs the same models quietly in a closet, as our Mac mini guide for local LLMs explains, and our Claude Code and local LLM setup on a Mac mini shows what that looks like in practice.

Our sources

We have not tested an RTX Spark laptop ourselves yet. Everything measured in this guide comes from these reviews, which we trust because they publish their methods:

• c’t 3003: RTX Spark im Test, der MacBook-Killer? (local LLMs, memory, Linux)
• Hardware Canucks: RTX Spark vs the fastest MacBook Pro (CPU, Adobe, AI, battery, gaming)
• GeekaWhat: NVIDIA RTX Spark review (gaming, DaVinci Resolve)
• Linus Tech Tips: My first 15 minutes with RTX Spark (agents, software)
• Notebookcheck: ASUS ProArt P16 RTX Spark review
• NVIDIA: RTX Spark product page

When our own laptop arrives, we will add our own experience here.

FAQ

What is NVIDIA RTX Spark?

RTX Spark is NVIDIA’s laptop and small-desktop platform built on the N1X superchip: an Arm-based Grace CPU with up to 20 cores, a Blackwell RTX GPU with up to 6,144 cores and up to 128 GB of LPDDR5X unified memory that the CPU and GPU share. The laptops run Windows 11 on Arm.

Is an RTX Spark laptop good for local LLMs?

Yes, that is its main strength. Because the GPU can use most of the shared memory, models that do not fit into the 8 GB of a normal laptop GPU run fully on the GPU. In c’t 3003’s test a 27B model ran at about 14.5 tokens per second and a 35B mixture-of-experts model at about 61 tokens per second.

How much memory do I need in an RTX Spark laptop?

For local AI, 64 GB. c’t 3003 ran out of room with 32 GB as soon as a model, a browser and a creative app ran together and recommends at least 48 GB. Memory is soldered, so it cannot be upgraded later.

Is RTX Spark better than a MacBook Pro?

Only for CUDA and some AI workloads on Windows. Hardware Canucks found local LLM speed close to an M5 Pro, faster on some models and slower on others, but the MacBook Pro was clearly ahead in CPU performance, battery life and Adobe apps such as Photoshop and Lightroom.

Can you game on an RTX Spark laptop?

Many games run well, Cyberpunk 2077 reached 63 fps at 1080p Ultra in c’t 3003’s test, but x86 games run through emulation. Hardware Canucks had 3 of 10 games fail to start, and anti-cheat support still varies. For a pure gaming laptop, an x86 laptop with an RTX 50-series GPU is the safer buy.