Navigate this guide · 1,803 words
The essential signal
- NVIDIA's first PC processor went on sale October 16, 2026: a Blackwell-plus-Grace Arm superchip with up to 128GB of shared memory, in laptops from six makers starting at $2,599.
- First reviews agree on the shape of it: class-leading mobile graphics and strong local-AI performance, held back by 2024-era CPU cores and uneven gaming on Windows on Arm.
- It puts NVIDIA into direct competition with Intel, AMD and Qualcomm for the PC's central processor — the first serious challenge to the old order in years.
On sale now
Friday, October 16, 2026, was the day NVIDIA stopped being just the graphics company inside your laptop and became the company that made the whole computer. RTX Spark laptops went on sale around the world from six manufacturers — ASUS, Dell, HP, Lenovo, Microsoft and MSI — with prices starting at $2,599 [1][2][3]. Launch events ran in stores, including a Best Buy in San Francisco, in the choreographed style NVIDIA has perfected for its graphics launches [1].[1][2][3]
The pitch is a PC that runs AI agents and large models without the cloud. NVIDIA says these machines can run frontier AI models locally while a creative app keeps running alongside them — the kind of workload that used to mean a workstation or a data center connection [1][2]. Whether that promise survives contact with real reviewers is the story of the next few sections, because the first independent tests landed the same week as the hardware.[1][2][3]
What's inside the superchip
RTX Spark is a single package that puts two very different engines side by side: a Blackwell-generation RTX graphics chip and a Grace processor, an Arm-based CPU design NVIDIA built together with MediaTek [1][2]. The top configuration, called N1X, carries up to 20 CPU cores and 6,144 CUDA cores in the GPU — roughly the shader count of a desktop GeForce RTX 5070, according to launch coverage [1][2][3].[1][2][3]
The CPU cores themselves are Arm's Cortex-X925 and Cortex-A725 designs, which reviewers were quick to note are 2024-generation architecture — not the newest Arm cores available, a detail that matters later [4]. The GPU side carries fifth-generation Tensor Cores with support for FP4, a low-precision number format that trades some accuracy for much faster AI inference [3]. It is, in short, a data-center idea — NVIDIA's Grace Blackwell thinking — shrunk into a laptop power envelope.[1][2][3]
One pool of memory
The design decision everything else hangs on is the memory: up to 128GB of LPDDR5X shared in a single pool between the CPU and the GPU, instead of the usual split where the graphics chip gets its own smaller slice [1][2]. A large AI model has to fit in memory to run at all, and on most laptops the GPU's private memory is the bottleneck that forces models back to the cloud [1].[1][2]
Shared memory removes that wall. NVIDIA's argument, echoed across the launch coverage, is that developers no longer need to copy data between host and device memory, and that a laptop can hold models that previously needed a server [1][2]. WebProNews notes that a 74-billion-parameter Nemotron model, with 9 billion active parameters, can run in 24GB of memory on this architecture — the kind of figure that makes local AI concrete rather than theoretical [2]. Memory, more than the logo on the box, is what you are actually buying here.[1][2]
A petaflop in a laptop
NVIDIA puts the platform's AI performance at up to 1 petaflop, measured in FP4 precision — the low-precision format used to run AI models quickly [1][2]. That number is doing marketing work, as all such numbers do, but the underlying claim is specific: chatbots, coding assistants, image generators and autonomous agents running on the machine in front of you instead of in a distant data center [1][2].[1][2]
The reason this matters beyond bragging rights is latency and privacy. Every request that stays on the device is a request that never crosses the network, never waits on a server queue, and never hands your data to someone else's computer [2]. Business technology coverage frames it as a genuine new option for local AI workloads — developers, data scientists and creative teams who need on-device acceleration now have a Windows machine built around that job [2]. The petaflop is the headline; the architecture underneath is the story.[1][2]
The road to the shelves
This did not appear overnight. The machines were first shown publicly at Microsoft's Windows and Surface event on October 7, where Microsoft also laid out plans for Copilot to act on the files stored on your PC — the software half of the same bet [1]. Preorders and confirmed pricing from the partners opened the same week, giving retailers a short window to gauge demand before the hard October 16 release date [3].[1][3]
That sequencing — reveal, preorder, retail within about ten days — is deliberate. It kept RTX Spark in the news cycle through launch week instead of burning all the attention on a single announcement day [1][3]. Six partners had confirmed products with preorders open as of October 7: ASUS, Dell, HP, Lenovo, Microsoft and MSI, with ASUS, HP and Microsoft publishing exact pricing first [3].[1][3]
What it costs
The entry price is $2,599, which buys the base configuration [1][2]. From there the ladder climbs steeply, and it climbs on memory: ASUS's ProArt P16 with 128GB of unified memory and a 2TB SSD lists at $6,999.99, the most expensive confirmed configuration, sitting above Microsoft's $5,999.99 RTX Spark Dev Box with the same memory and storage [3].[1][2][3]
That spread tells you who this is for in its first generation. This is not a mainstream laptop play; it is a premium machine for people whose work justifies workstation money — developers running local models, creators working with enormous scenes, researchers who want a cluster node that fits in a backpack [1][2]. The 128GB configuration is where the capability lives, and it is also where the invoice lives.[1][2][3]
The first-wave lineup
Six manufacturers, one chip, a spread of designs. The confirmed names so far include ASUS's ProArt P16, Microsoft's Surface Laptop Ultra and RTX Spark Dev Box, and HP's OmniBook Ultra 16, while Dell, Lenovo and MSI have products confirmed without published pricing yet [3]. All of them run Windows 11 on Arm, with Microsoft's execution containers as part of the software story [3].[1][3]
What is striking is how quickly the lineup filled out. Having six major OEMs shipping on day one is not how experimental platforms usually launch — it signals that the PC industry sees real demand for local-AI machines, or at least real fear of missing one [1][3]. Acer and Gigabyte, two names from earlier reporting, had not confirmed products as of launch week, which suggests the first wave is a deliberate coalition rather than an industry-wide stampede [3].[1][3]
Reviews: the graphics lead
The first independent benchmarks arrived with the hardware, and on graphics they are emphatic. In 3DMark Time Spy Extreme, the N1X scored 6,060 — against 4,689 for AMD's Ryzen AI Max+ 395 and 3,226 for Intel's top mobile chip [4]. Steel Nomad came in at 2,912 versus 1,895 for AMD and 1,483 for Intel, and the Solar Bay ray-tracing test at 10,594 against 6,047 and 4,015 [4][6].[4][6]
Those are synthetic tests, but the gap is large enough to survive skepticism about benchmarks. Tech Times' review roundup calls it class-leading mobile GPU performance without qualification, noting that the N1X's larger compute block and unified memory produce clear advantages even though discrete laptop GPUs still enjoy higher raw memory bandwidth [4]. For creative work — 3D rendering, video, GPU compute — this is the strongest integrated graphics ever shipped in a Windows laptop, and it is not close.[4][6]
Reviews: AI muscle
The AI benchmarks tell a similar story. Tom's Hardware's testing, summarized by TweakTown, found local AI running a 27-billion-parameter Qwen model at roughly twice the speed of AMD's Ryzen AI Halo device, and notably faster than Apple's MacBook Pro with the M5 Max chip [5]. With lighter models the lead held, which is encouraging for anyone eyeing these machines for on-the-go AI work [5].[4][5]
This is the workload RTX Spark was built for, so the result matters more than any game frame rate. Running capable models locally — fast, private, offline — is the entire justification for the unified memory and the Tensor Cores [4][5]. The reviewers' consensus is that NVIDIA's CUDA software stack, carried over from its data-center dominance, is a genuine moat here: the tools developers already use run natively, while competitors are still building theirs [5].[4][5]
Reviews: the CPU shows its age
Here the story turns. Those Cortex-X925 and A725 cores are Arm's 2024 designs, and in single-threaded tests the age shows: Geekbench 7 single-core scores around 2,604–2,710, competitive with Snapdragon X Elite territory but trailing Apple's M5 Pro at 3,690 [4][6]. Multi-core throughput is genuinely competitive with today's best mobile platforms — code compilation, video encoding and multi-threaded work are all strong — but single-threaded tasks like browser performance and game logic trail Apple's current generation by a meaningful margin [4].[4][6]
Battery life splits by workload too. HotHardware measured 14 hours and 35 minutes of video playback on a Surface Laptop Ultra, with multi-threaded performance virtually unchanged on battery — confirming NVIDIA's claim that the machine holds full performance unplugged [4]. But Tom's Guide managed about 10 hours and 15 minutes in a web-browsing test against 21:25 for a 16-inch MacBook Pro, so all-day claims depend heavily on what "all day" means for you [6]. The practical summary from the reviews: a superb engine with an aging transmission.[4][6]
Gaming: playable, with caveats
Gaming is where Windows on Arm's immaturity shows. Trusted Reviews measured Cyberpunk 2077 at 1080p Ultra averaging nearly 70 frames per second — genuinely impressive for the power envelope — but heavy ray tracing pushed closer to 30 fps, and NotebookCheck's testing found the chip notably faster than AMD's fastest integrated GPU while still trailing discrete graphics [4][5]. Tom's Hardware concluded that the 300 GB/s memory bandwidth becomes the visible constraint in CPU-heavy titles [5].[3][4][5]
The deeper problem is compatibility, not speed. Game compatibility on Windows on Arm has historically been inconsistent, particularly for titles with strict anti-cheat software, and that has not changed at launch [3]. Some anti-cheat systems were simply not yet compatible with Arm in reviewers' testing [5]. The honest guidance from the coverage: creators and developers who also game will be happy; anyone buying primarily to play should buy a traditional gaming laptop instead [4][5].[3][4][5]
Why it matters anyway
Step back from the benchmarks and the significance is structural. For decades, the processor and the graphics chip in a Windows laptop came from different vendors; RTX Spark collapses both into a single NVIDIA part, giving the company a slice of the laptop bill-of-materials that used to belong entirely to CPU makers [2][3]. It puts NVIDIA in direct competition with Intel and AMD on x86's home turf, and with Qualcomm as the incumbent Windows-on-Arm vendor — Qualcomm's Snapdragon X2 Elite already ships broadly, so RTX Spark enters as a premium alternative rather than a replacement, at least in this generation [2][3].[2][3]
It lands at an awkward moment for everyone else. Intel and AMD have spent two years marketing NPU-equipped chips for on-device AI without matching anything like this memory headroom, and neither has an answer to 6,144 CUDA cores in a thin laptop [2]. First-generation caveats apply — premium pricing, Arm software friction, unproven battery claims in real mixed use — but the direction is unmistakable. The PC's central processor is contested ground again, for the first time in years [2][3].[2][3]
Questions, answered.
What exactly is NVIDIA RTX Spark?
NVIDIA's first PC processor: a single "superchip" combining a Blackwell RTX GPU with an Arm-based Grace CPU (co-designed with MediaTek), up to 20 CPU cores and 6,144 CUDA cores, with as much as 128GB of memory shared between them [1][2].[1][2]
How much do RTX Spark laptops cost?
From $2,599 at the entry level to $6,999.99 for the top ASUS ProArt P16 configuration with 128GB of memory and 2TB of storage. Memory is what drives the price up [1][2][3].[1][2][3]
Will my Windows apps and games run on it?
Native Arm apps run directly; most x86 apps run through Windows' emulation layer. Game compatibility is the weak spot — some titles with strict anti-cheat software do not work yet [3][5].[3][5]
Read the sources.
- Nvidia RTX Spark laptops on sale: models, prices and local AIMadRobot · Accessed 2026-10-22
- NVIDIA's RTX Spark Laptops Mark Bold Push Into Arm-Powered Windows AI HardwareWebProNews · Accessed 2026-10-22
- NVIDIA RTX Spark Price: $2,599 to $5,999 (Oct. 16) [2026]Tech Insider · Accessed 2026-10-22
- RTX Spark Reviews Are In: Class-Leading GPU Meets Aging CPU Cores at LaunchTech Times · Accessed 2026-10-22
- Early RTX Spark reviews show impressive AI performance while gaming falls a little shortTweakTown · Accessed 2026-10-22
- RTX Spark Laptops Are Out: Prices, Models and What the First Reviews SayTeckhustlers · Accessed 2026-10-22
Contributor articles and source links do not guarantee error-free reporting. See our editorial policy.
