NVIDIA has confirmed that its long-rumored RTX Spark PCs will reach store shelves in October 2026, and the chip powering them, codenamed N1X, is arriving in two distinctly different configurations rather than one, a follow-up to the initial October launch confirmation. According to reports from Wccftech, NotebookCheck, and MacDailyNews, the flagship N1X pairs a 20-core CPU with a 6,144-core GPU and up to 128GB of unified memory, while a second, cheaper variant ships with an 18-core CPU, a 5,120-core GPU, and a hard ceiling of 32GB. That is a four-times gap in maximum memory between two chips carrying the same N1X name, and it is the detail that matters most for anyone trying to figure out which RTX Spark machine to actually buy.
The split is not a minor SKU footnote. It is NVIDIA’s first attempt to sell a complete Arm-based Windows PC platform at scale, and the company chose to do it by drawing a hard line between a machine built for serious local AI work and one built to look like it. Nvidia rtx spark buyers now face a decision that looks a lot like choosing between a workstation and a laptop wearing a workstation’s badge, and the specs gap explains why.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
NVIDIA Confirms Two N1X Configurations for RTX Spark PCs
The core news is straightforward: RTX Spark PCs built around the N1X chip are arriving next month in two configurations, confirmed across multiple outlets including Wccftech, NotebookCheck, and MacDailyNews. The higher-end variant carries a 20-core CPU, a 6,144-core GPU, and unified memory scaling up to 128GB. The second variant drops to an 18-core CPU, a 5,120-core GPU, and caps out at 32GB of unified memory. Both chips share the same Grace-Blackwell design lineage and the same N1X branding, which is exactly why the gap between them is confusing shoppers already searching for nvidia rtx spark specs comparisons ahead of launch.
Tom’s Hardware, which covered NVIDIA’s original Computex reveal of the RTX Spark superchip, described the platform as an attempt to turn Windows into what NVIDIA calls an “agentic AI OS,” combining an Arm CPU with a Blackwell GPU and up to 128GB of unified memory in a single package. That description matches only the flagship configuration. The second-tier chip, aimed at thinner and cheaper laptops, was not part of the original splashy pitch and has emerged more quietly through spec leaks and OEM roadmap reporting over the summer.
What is notable is how little daylight NVIDIA left between announcement and shipping date. The RTX Spark superchip was first shown publicly at NVIDIA’s Computex keynote in the early summer of 2026, and reports now place actual retail availability in October, a turnaround of only a few months for a brand-new CPU-and-GPU platform entering a market NVIDIA has never directly competed in before.
Inside the Flagship N1X: 20 Cores, 6,144 CUDA Cores, 128GB Memory
The top-tier N1X configuration is the chip NVIDIA actually wants reviewers writing about. It pairs a 20-core Grace-derived CPU with a Blackwell RTX GPU carrying 6,144 CUDA cores, a core count that multiple outlets have pointed out lines up with a desktop RTX 5070. That is a striking comparison for a chip destined for laptops and compact desktops rather than a discrete graphics card slot, and it is the number NVIDIA is leaning on to justify calling RTX Spark a genuine AI workstation rather than an AI-branded ultrabook.
Memory is where the flagship chip separates itself hardest from the rest of the Windows PC market. Reports describe unified LPDDR5X memory scaling up to 128GB, shared between the CPU and GPU rather than split into separate system RAM and VRAM pools. That unified pool, connected via NVIDIA’s NVLink-C2C chip-to-chip interconnect, is the same basic architectural idea behind NVIDIA’s data-center Grace Blackwell superchips, shrunk down into a form factor that can sit on a desk or ride in a backpack.
NVIDIA has not published an official memory bandwidth figure for N1X, and outlets covering the chip have been careful to flag that gap rather than guess at a number. What is confirmed is the process node: N1X is built on a 3-nanometer node through NVIDIA’s collaboration with MediaTek, the same partnership that produced the CPU core design underpinning the Grace side of the chip. The flagship configuration is also the only one of the two confirmed for both laptops and compact mini PCs at launch, according to Wccftech’s reporting, which matters for anyone hoping to buy RTX Spark as a small-form-factor desktop rather than a laptop.
The Second Tier: 18 Cores, 5,120 CUDA Cores, 32GB Ceiling
The lower configuration is a meaningfully smaller chip, not just a clock-speed-limited version of the flagship. It carries an 18-core CPU, two fewer cores than the top model, and a GPU with 5,120 CUDA cores, roughly 17% fewer shader units than the 6,144-core flagship. The bigger cut lands on memory: the second-tier N1X tops out at 32GB, a quarter of the flagship’s ceiling.
That 32GB ceiling is the number that should give AI-curious buyers pause. Local large language model inference is memory-bound before it is anything else, and a 32GB unified pool shared between the OS, applications, and any model weights loaded for inference leaves far less headroom than the flagship’s 128GB. Reports indicate the 18-core, 5,120-core configuration is laptop-only at launch, with no mini PC or compact desktop variant confirmed, which reinforces the read that NVIDIA is positioning it as the “RTX Spark, but affordable” option rather than a true peer to the flagship.
Both chips reportedly share the same 3nm process node and the same fundamental Grace-Blackwell architecture, meaning the split is a binning and configuration decision rather than two fundamentally different pieces of silicon. That is a common practice in CPU and GPU manufacturing, but rarely has the gap between a flagship and its cut-down sibling been this wide within a single, freshly launched product line.
Why NVIDIA Split the Spark Lineup Into Two SKUs
NVIDIA rarely enters a market with a single product when it can build a ladder instead, and RTX Spark follows the same playbook the company has used for years in discrete GPUs. Splitting N1X into two configurations lets NVIDIA and its OEM partners hit two different price points and two different buyer intentions with one launch instead of two separate development cycles.
The flagship, with its 128GB ceiling, is clearly aimed at developers, researchers, and small studios who want to run meaningfully large local models, fine-tune smaller ones, or handle heavy generative AI workloads without paying for cloud GPU time. The 32GB variant is aimed at a much broader audience: buyers who want the RTX Spark badge, Windows AI features, and better-than-integrated graphics performance, but who were never going to load a 70-billion-parameter model onto their laptop anyway.
This segmentation also protects NVIDIA’s pricing on the high end. If every RTX Spark shipped with 128GB standard, NVIDIA would either have to eat the cost of that memory across a mainstream price point or price the entire line out of reach of casual buyers. By reserving the 128GB configuration for a distinct, presumably pricier tier, NVIDIA can chase both the enthusiast AI-PC crowd and the mainstream premium-laptop buyer without compromising either segment’s economics. No official pricing has been disclosed for either configuration as of this reporting, and outlets covering the launch describe Spark systems only in general terms as a premium, high-end category.
RTX Spark N1X Spec Comparison
| Spec | N1X Flagship | N1X Second Tier |
|---|---|---|
| CPU cores | 20-core Grace-based CPU | 18-core Grace-based CPU |
| GPU cores | 6,144 Blackwell RTX CUDA cores | 5,120 Blackwell RTX CUDA cores |
| Unified memory | Up to 128GB LPDDR5X | Up to 32GB LPDDR5X |
| Process node | 3nm (NVIDIA/MediaTek) | 3nm (NVIDIA/MediaTek) |
| Interconnect | NVLink-C2C | NVLink-C2C |
| GPU core-count peer | Roughly desktop RTX 5070 class | Below RTX 5070 class |
| Confirmed form factors | Laptops and mini PCs/desktops | Laptops only |
| Launch window | October 2026 | October 2026 |
| Official pricing | Not yet disclosed | Not yet disclosed |
Grace Blackwell Architecture, Explained for PC Buyers
N1X did not appear out of nowhere. It is a client-market descendant of NVIDIA’s Grace Blackwell data-center superchips, the same architectural family behind NVIDIA’s server-grade AI infrastructure. In the data center, Grace Blackwell pairs an Arm-based Grace CPU with a Blackwell GPU over a high-speed NVLink-C2C link, letting both processors share a single pool of memory instead of shuttling data back and forth over a slower bus. NVIDIA’s own Grace CPU Superchip documentation describes that unified-memory approach as central to the platform’s efficiency for large AI workloads.
N1X shrinks that same idea down into a chip that fits inside a laptop chassis or a mini PC case. Both configurations include NVIDIA’s fifth-generation Tensor Cores with FP4 support, a lower-precision numeric format that trades some accuracy for significantly higher throughput on inference workloads. The Register, covering NVIDIA’s original announcement, cited figures as high as 500 TFLOPS of dense FP4 performance and roughly 1 PFLOP with sparsity enabled on the flagship configuration, framing RTX Spark as bringing data-center-style AI throughput down into a consumer or prosumer machine for the first time.
It is worth being precise about what N1X is not. It is not a rebadged GB200 or any other explicitly named data-center part, and none of the reporting around RTX Spark has mapped N1X directly onto NVIDIA’s internal GB10 or GB20 naming scheme. N1X appears to be a distinct, PC-market-specific brand sitting on top of the same underlying Grace Blackwell design philosophy, not a repackaged server chip.
Which OEMs Are Building RTX Spark Machines
Six PC manufacturers are consistently named across coverage of the RTX Spark launch: ASUS, Dell, HP, Lenovo, MSI, and Microsoft, the last of which is expected to ship Surface-branded RTX Spark hardware. That is a broad launch-day partner list for a first-generation NVIDIA PC platform, and it signals that major OEMs see enough demand for an Arm-based, AI-focused Windows machine to commit engineering resources before a single unit has shipped.
Reporting has not yet broken down which specific OEM will ship which N1X configuration, or whether any single manufacturer plans to offer both the 128GB flagship and the 32GB second tier under one product line. What has been confirmed is that the flagship chip’s mini-PC and compact-desktop suitability suggests at least some of the six named partners are building small-form-factor desktops rather than laptops exclusively, a form factor NVIDIA has effectively never sold a first-party Windows platform into before. ASUS has already reportedly adjusted its RTX Spark stock orders heading into the fourth quarter, while ASUS and MSI’s existing RTX Spark laptops reportedly sold out shortly after their initial listings.
Local AI Inference: What 4x More Memory Actually Buys You
For the software engineers and AI developers who are the most likely early adopters of RTX Spark, memory capacity is the single most consequential number on the spec sheet. Model weights have to fit in memory before a single token of inference happens, and larger unified memory pools let a machine hold bigger models, longer context windows, and more simultaneous processes without falling back to slow disk swapping or aggressive quantization.
The 128GB flagship configuration puts RTX Spark in a category that, until now, has mostly belonged to Apple’s higher-end Mac Studio configurations and a handful of specialized workstation cards. A 32GB ceiling on the second-tier chip, by contrast, keeps that machine firmly in mid-range territory: workable for smaller open-weight models and lighter fine-tuning jobs, but not the machine a developer would reach for to run the largest openly available models locally.
Developers checking whether a given RTX Spark configuration can handle a specific model will likely rely on the same kind of system inspection tools already common in NVIDIA’s CUDA ecosystem. A simple driver query, for instance, remains the fastest way to confirm how much unified memory a system is actually reporting to the GPU stack:
nvidia-smi --query-gpu=name,memory.total,memory.used --format=csvThat kind of check will matter more on RTX Spark than on a typical discrete-GPU Windows laptop, precisely because the CPU and GPU are drawing from the same unified pool rather than separate system RAM and VRAM allocations.
RTX Spark N1X vs Apple Mac Studio vs AMD Strix Halo
RTX Spark is not launching into an empty field. Apple’s Mac Studio lineup has offered unified-memory configurations reaching well past 100GB for several product generations, built around Apple’s own M-series silicon and tightly integrated with macOS and Apple’s Metal and Core ML frameworks rather than CUDA. AMD’s Strix Halo platform takes a different approach again, pairing an integrated RDNA GPU with AMD’s XDNA AI accelerator inside a conventional x86 laptop chip, targeting many of the same AI-PC buyers without offering a CUDA-compatible software stack.
| Platform | Max Unified Memory | CPU Architecture | AI Software Ecosystem |
|---|---|---|---|
| NVIDIA RTX Spark N1X (flagship) | 128GB LPDDR5X | Arm (Grace-derived) | Full CUDA stack, FP4 Tensor Cores |
| NVIDIA RTX Spark N1X (second tier) | 32GB LPDDR5X | Arm (Grace-derived) | Full CUDA stack, FP4 Tensor Cores |
| Apple Mac Studio (Apple Silicon) | Configurable past 100GB | Apple Silicon (Arm) | Metal, Core ML, no CUDA |
| AMD Strix Halo | Shared system DDR5/LPDDR5 | x86 (Zen-based) | ROCm, no CUDA |
The practical difference for developers comes down to software gravity as much as raw specs. CUDA remains the default target for the overwhelming majority of open-source machine learning tooling, from PyTorch to the inference frameworks built around today’s open-weight model releases. That gives RTX Spark’s flagship configuration a real advantage for developers who want to run existing, unmodified AI tooling locally on Windows rather than adapting workflows to Metal or ROCm. Apple’s advantage remains in its own tightly integrated ecosystem and its long track record of shipping large unified-memory configurations at scale, while AMD’s Strix Halo competes primarily on x86 compatibility and price rather than raw AI throughput.
Market Impact: How Intel, AMD, and Apple Should Read This
NVIDIA entering the PC chip market as a full CPU-and-GPU vendor, rather than purely a component supplier selling discrete graphics cards into other companies’ systems, is a structural shift for the Windows PC industry. For decades, Windows laptops and desktops have been built around x86 processors from Intel or AMD, with NVIDIA supplying add-in GPUs. RTX Spark collapses that division for at least one category of premium AI-focused machine, putting NVIDIA in direct competition with the CPU vendors it has spent years selling alongside.
The two-tier N1X split also puts pressure on how Intel and AMD think about their own AI-PC roadmaps. If NVIDIA can offer a 128GB unified-memory machine with a full CUDA software stack at the top of its range, and a genuinely cheaper 32GB option at the bottom, Intel’s and AMD’s answer needs to cover both ends of that range too, not just compete at a single price point. AMD’s Strix Halo family already attempts exactly that kind of range-covering strategy, but without CUDA compatibility, it is negotiating from a software position, not a hardware one.
For Apple, the competitive pressure is more indirect. RTX Spark is a Windows platform, not a macOS one, so it does not directly compete for the same purchase decision as a Mac Studio. But it does compete for the same developer mindshare: engineers deciding which unified-memory machine to buy for local AI experimentation now have a Windows-native, CUDA-compatible option with the same 128GB memory ceiling Apple has offered for several product cycles. NVIDIA’s broader push toward local AI hardware over cloud subscriptions has already rattled chip stocks once this year, with Intel and AMD shares dropping after NVIDIA’s initial PC-chip market entry.
Historical Context: NVIDIA’s Long Road From GPUs to Full PCs
NVIDIA has flirted with owning more of the PC stack before. The company built reference-design motherboards in the early 2000s under its nForce chipset brand, and it later tried to break into mobile and PC processors with the Tegra line, which found its biggest success not in phones or Windows PCs but inside the Nintendo Switch. NVIDIA also shipped Arm-based server CPUs under the Grace name well before RTX Spark, establishing the CPU design lineage that N1X now draws on.
What is different this time is timing and motivation. The current AI boom has made unified CPU-GPU memory architectures, once a niche data-center concept, into a mainstream selling point that ordinary PC buyers are beginning to search for and understand. NVIDIA’s Grace Blackwell platform already dominates AI data-center infrastructure, and RTX Spark represents the company’s attempt to extend that same architecture, and the developer familiarity with CUDA that comes with it, all the way down to a desk or a backpack. It is a far more direct assault on the traditional x86 PC market than Tegra ever attempted, precisely because it arrives at a moment when “how much can I run locally” has become a genuinely mainstream question for developers.
What Remains Unconfirmed Ahead of the October Launch
Several important details have not yet been made public, and readers should treat any number outside of the confirmed core counts and memory ceilings with caution. NVIDIA has not disclosed an exact memory bandwidth figure in gigabytes per second for either N1X configuration, nor has it published official thermal design power ratings for shipping systems. Pricing for both the flagship and second-tier configurations remains undisclosed, and no outlet has yet reported a specific calendar release date beyond the general “October 2026” window repeated across coverage.
Per-OEM product mapping is also still murky. While ASUS, Dell, HP, Lenovo, MSI, and Microsoft are all confirmed as launch partners, none of the six has published a specific model name, chassis design, or confirmed which of the two N1X configurations its first machines will use. Buyers hoping to pre-order a specific RTX Spark mini PC or laptop configuration will likely need to wait for individual OEM announcements closer to the October launch window.
What to Watch: 5 Predictions for RTX Spark’s Launch
- Expect the 128GB flagship configuration to command a significant price premium over the 32GB second tier, likely positioning it closer to high-end creator or workstation laptop pricing than mainstream ultrabook pricing.
- OEMs will likely differentiate primarily on chassis and cooling rather than internal spec changes, since both N1X variants appear to be fixed configurations rather than customizable by manufacturer.
- Independent memory bandwidth benchmarks will surface within days of the October launch, since reviewers will almost certainly run standard tools the moment retail units ship, filling in the one major spec NVIDIA has left undisclosed.
- Developer community interest will concentrate heavily on the 128GB flagship, given that the 32GB second tier offers little advantage over existing high-end integrated-GPU laptops for local AI workloads specifically.
- Intel and AMD are likely to accelerate their own unified-memory AI-PC roadmap announcements in the weeks following RTX Spark’s launch, using their existing x86 compatibility as the primary counter-argument to NVIDIA’s CUDA advantage.
Should Developers Wait for RTX Spark or Buy Now?
For developers currently weighing a new machine purchase for local AI work, the calculus depends heavily on which N1X configuration ends up in reach financially once pricing is announced. The 128GB flagship, assuming it carries the premium pricing its specs suggest, competes directly with high-end Mac Studio configurations and specialized AI workstations rather than typical premium laptops. The 32GB second tier is a much harder sell for anyone buying specifically for AI workloads, since 32GB of shared memory, split between the operating system and any loaded models, leaves comparatively little room for anything beyond smaller open-weight models.
Buyers who do not need local AI inference at all, and simply want a fast, modern Windows machine with strong Tensor Core-accelerated creative and productivity features, may find the 32GB configuration perfectly adequate once pricing lands. The distinction NVIDIA has drawn between its two N1X chips is really a distinction between two different customers, not two versions of the same product. For a broader look at where AI chip pricing and availability are headed across the industry, see our AI chips 2026 coverage.
Frequently Asked Questions
What is the NVIDIA N1X chip?
N1X is NVIDIA’s codename for the system-on-chip powering its RTX Spark PC platform, combining a Grace-derived Arm CPU with a Blackwell RTX GPU and unified LPDDR5X memory, built on a 3nm process through NVIDIA’s collaboration with MediaTek.
What are the two RTX Spark N1X configurations?
The flagship configuration pairs a 20-core CPU with a 6,144-core GPU and up to 128GB of unified memory. The second-tier configuration has an 18-core CPU, a 5,120-core GPU, and a 32GB memory ceiling, and is currently confirmed for laptops only.
When do RTX Spark PCs launch?
Reports from Wccftech, NotebookCheck, and MacDailyNews place the launch window in October 2026, though no outlet has published an exact calendar date yet.
How much will RTX Spark PCs cost?
NVIDIA has not disclosed official pricing for either N1X configuration. Coverage of the platform describes it only in general terms as a premium, high-end category comparable to existing high-end gaming or creator laptops.
Which companies are making RTX Spark PCs?
ASUS, Dell, HP, Lenovo, MSI, and Microsoft are the six confirmed launch partners, with Microsoft expected to ship Surface-branded RTX Spark hardware.
Can RTX Spark PCs run large language models locally?
The 128GB flagship configuration is built specifically for that use case, with enough unified memory to hold significantly larger models than typical Windows laptops. The 32GB second-tier configuration is better suited to smaller models and lighter AI workloads.
How does RTX Spark compare to Apple’s Mac Studio for AI work?
Apple’s Mac Studio has offered unified-memory configurations past 100GB for several generations, built around Apple Silicon and macOS. RTX Spark’s flagship configuration matches that memory ceiling at 128GB while offering full CUDA compatibility, which is the dominant software stack for open-source AI tooling.
Is N1X the same as NVIDIA’s data-center GB10 or GB20 chips?
No. N1X shares the same Grace Blackwell architectural approach as NVIDIA’s data-center superchips, but reporting on the platform has not mapped it directly onto NVIDIA’s GB10 or GB20 naming, and N1X appears to be a distinct, PC-market-specific design.
Related Coverage
- Nvidia RTX Spark N1X Launches October: 128GB RAM [2026]
- NVIDIA Confirms DLSS 5 for RTX 40 GPUs, RTX 50 First [2026]
- Intel Nova Lake Release Date Leak: Q1 2027, 52 Cores [2026]
- DDR5 RAM Hits $400 for 32GB as Prices Soar 485% [2026]
- How to Overclock a GPU: RTX 5080/5090 & RX 9070 XT in 12 Steps [2026]
