NVIDIA RTX Spark
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 2,344 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 2,344 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA RTX Spark is a consumer "superchip" for Windows personal computers, announced by NVIDIA on May 31, 2026, at the company's GTC Taipei keynote held during COMPUTEX 2026. Developed in collaboration with Microsoft, RTX Spark pairs a 20-core NVIDIA Grace Arm CPU with an NVIDIA Blackwell RTX GPU on a single package, joined by a coherent NVLink-C2C interconnect and up to 128GB of unified memory. NVIDIA positions it as the foundation for a new class of Windows laptops and compact desktops built for on-device "personal AI agents," and as the client-platform sibling of the developer-focused NVIDIA DGX Spark. The chip is the productized version of the part previously rumored under the codename "N1X," its Arm CPU co-designed with MediaTek, and it is the first member of a multi-generation NVIDIA roadmap for the PC. RTX Spark laptops and desktops are scheduled to ship in the fall of 2026 from a broad set of original-equipment manufacturers.[1][2][3]
| Attribute | Detail |
|---|---|
| Product name | NVIDIA RTX Spark (superchip / platform) |
| Type | Arm-based CPU+GPU system-on-package ("superchip") for PCs |
| Announced | May 31, 2026, GTC Taipei keynote at COMPUTEX 2026 (Taipei) |
| Developed with | Microsoft (Windows); Arm CPU co-designed with MediaTek |
| Reported codename | N1X (full-spec part) and a leaner N1 variant |
| CPU | 20-core NVIDIA Grace (Arm) |
| GPU | NVIDIA Blackwell RTX, 6,144 CUDA cores, 5th-generation Tensor Cores with FP4 |
| Interconnect | NVIDIA NVLink-C2C (CPU-to-GPU) |
| Memory | Up to 128GB unified LPDDR5X |
| AI performance | 1 petaflop (NVIDIA figure, FP4) |
| Process node | TSMC 3nm |
| Availability | Fall 2026 (laptops and compact desktops) |
| Launch OEMs | ASUS, Dell, HP, Lenovo, Microsoft Surface, MSI; Acer and GIGABYTE to follow |
| Relationship | Client variant of the same Grace Blackwell family as DGX Spark / GB10 |
RTX Spark is an integrated CPU-plus-GPU package, what NVIDIA calls a "superchip," intended to power mainstream Windows computers rather than data-center accelerators. The company frames it as the first ground-up reengineering of the consumer PC in decades, built around the idea that people will increasingly interact with their computers through AI agents that run locally on the device instead of in the cloud. In NVIDIA's framing, the same hardware should be capable of running large AI models on battery power, while still handling demanding creative and gaming workloads.[1][2]
NVIDIA chief executive Jensen Huang (Jensen Huang) tied the launch directly to that shift in how PCs are used. "The PC is being reinvented," he said in the announcement. "For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask, and the PC does the work."[2]
The platform bundles much of NVIDIA's accumulated PC technology stack, including CUDA, NVIDIA RTX, DLSS, FP4 low-precision math, TensorRT, OptiX, Reflex and G-SYNC, into a single Arm-based design aimed at slim laptops with all-day battery life and small, power-efficient desktops.[2]
NVIDIA unveiled RTX Spark during Jensen Huang's keynote at GTC Taipei, the company's developer conference co-located with COMPUTEX 2026 in Taipei, on May 31, 2026. Some press coverage dates the reveal to June 1, 2026, reflecting local time zones and the COMPUTEX show schedule, but the event is the same keynote.[1][2][4]
The reveal ended a long stretch of speculation about an NVIDIA-designed Windows-on-Arm processor that had circulated under the codename "N1X." Press reporting indicated that the part was co-designed with MediaTek and that RTX Spark is the consumer-facing product built on it.[3][4]
At the heart of RTX Spark is a tightly coupled pairing of an NVIDIA Grace Arm CPU and an NVIDIA Blackwell RTX GPU on the same package, a client-scale expression of the Grace Blackwell architecture that NVIDIA uses across its product lines. The CPU and GPU are linked by NVIDIA's NVLink-C2C (NVLink) chip-to-chip interconnect, which provides a high-bandwidth, cache-coherent path between the two so they can share a single pool of memory rather than copying data back and forth across a slower bus.[1][5]
That shared pool is the defining feature for AI work. RTX Spark offers up to 128GB of unified LPDDR5X memory addressable by both the CPU and GPU, which is what allows comparatively large models to be loaded entirely on-device. Press analysis put the full-spec part on TSMC's 3nm process, with one detailed account citing roughly 70 billion transistors, around 300 GB/s of memory bandwidth, and an NVLink-C2C link running at about 600 GB/s. NVIDIA's headline performance figure for the platform is 1 petaflop of AI compute, expressed in the FP4 precision supported by the fifth-generation Tensor Cores.[1][5][4]
| Component | Specification | Source attribution |
|---|---|---|
| CPU | 20-core NVIDIA Grace (Arm), co-designed with MediaTek | NVIDIA |
| GPU | NVIDIA Blackwell RTX, 6,144 CUDA cores | NVIDIA |
| Tensor Cores | 5th generation, FP4 precision | NVIDIA |
| CPU-GPU interconnect | NVIDIA NVLink-C2C (~600 GB/s reported) | NVIDIA / press |
| Unified memory | Up to 128GB LPDDR5X | NVIDIA |
| Memory bandwidth | ~300 GB/s (reported) | Press (StorageReview) |
| AI performance | 1 petaflop (FP4) | NVIDIA |
| Process node | TSMC 3nm | Press (TrendForce) |
| Transistors | ~70 billion (reported) | Press (StorageReview) |
| Power envelope | Single-digit watts up to ~80W (reported) | Press (StorageReview) |
The figures above combine NVIDIA's own published specifications with details reported by trade press. Performance numbers such as the 1 petaflop AI figure are NVIDIA's claims and are stated at FP4 precision, which roughly quadruples the headline throughput compared with higher-precision formats. Independent benchmarks were not available at announcement.
NVIDIA's central pitch for RTX Spark is the ability to run sizable generative AI models locally. The company says a full-spec system can host large language models of up to 120 billion parameters and operate them with a context window of up to 1 million tokens, with the large unified memory pool and FP4 Tensor Cores doing the heavy lifting. These are NVIDIA performance claims tied to the top configuration.[1][5]
The platform leans heavily on NVIDIA's FP4 software work to make those models fit and run efficiently. NVIDIA has been shipping NVFP4 quantized checkpoints that compress models into the four-bit format the Tensor Cores accelerate, and the same optimizations that benefit DGX Spark carry over to RTX Spark because the two share the underlying Grace Blackwell design and unified-memory model.[6]
Beyond AI, NVIDIA highlighted creative and gaming performance as evidence that a single thin-and-light design can serve multiple audiences. The company cited the ability to render a 90GB-plus 3D scene with OptiX and DLSS, edit 12K 4:2:2 video using the Blackwell decoder, generate 4K AI video with RTX Video frame generation, and play graphically demanding titles at 1440p and over 100 frames per second, all without an external power supply. Trade coverage characterized the graphics performance as roughly in the class of a GeForce RTX 5070 laptop GPU.[2][4]
RTX Spark was announced jointly with Microsoft, and the Windows software story is as central to the launch as the silicon. The two companies described a native Windows experience for "personal agents," in which AI agents can act on the user's behalf across applications. To make that safe, Microsoft is adding new Windows security primitives covering agent identity, containment, policy and end-to-end security, and NVIDIA is contributing a runtime called NVIDIA OpenShell that packages agents to run securely on a person's primary device under user-defined policies. NVIDIA described OpenShell as an easy-to-deploy package for secure, on-device agents built on Microsoft's new agent security primitives.[1][2][6]
This is the practical meaning behind Huang's framing that users will "ask" the PC to do work rather than manually launching and operating applications. The agentic approach is part of the broader industry move toward agentic AI, software that can plan and carry out multi-step tasks, and NVIDIA's argument is that running such agents locally improves latency, privacy and cost compared with sending every request to the cloud.[2]
NVIDIA also reported broad software-ecosystem support, citing more than 100 Windows software providers adopting RTX Spark. Adobe is rearchitecting Photoshop and Premiere for the platform, with NVIDIA citing up to 2x faster AI and graphics performance, and other named adopters span creative tools such as Blackmagic Design, Blender, ComfyUI and OTOY as well as game studios including KRAFTON, NetEase, Remedy Entertainment and Riot Games.[2]
RTX Spark is closely related to NVIDIA's DGX Spark, the compact "personal AI supercomputer" NVIDIA introduced earlier built on the GB10 Grace Blackwell Superchip. Both products are built from the same Grace Blackwell family and both expose up to 128GB of unified LPDDR5X memory, and trade reporting noted that the full-spec RTX Spark mirrors the DGX Spark silicon very closely. The key difference is the target: DGX Spark is a Linux machine aimed at AI developers and researchers, while RTX Spark is a Windows-first consumer and creator product distributed through mainstream PC makers.[3][4][5]
Reporting on the chip family describes the lineup in terms of an N1X full-spec part and a leaner N1 variant, with the higher-end silicon shared between the DGX Spark developer device and the RTX Spark client platform and lower-power or smaller-memory configurations serving thinner laptops. Press accounts placed entry-level RTX Spark SKUs as low as 16GB of memory, scaling up to the 128GB flagship.[3][5]
At the same keynote NVIDIA also announced DGX Station for Windows, a far larger deskside system for enterprise developers built on the GB300 Grace Blackwell Ultra Desktop Superchip, with up to 748GB of coherent memory and up to 20 petaflops of FP4 performance, capable of running trillion-parameter models locally. That sits well above RTX Spark and is a different class of product, related to NVIDIA's data-center Grace Blackwell systems such as the GB300 NVL72 rack platform rather than to mainstream PCs.[7][2]
NVIDIA said RTX Spark laptops and compact desktops will be available in the fall of 2026. The launch lineup includes designs from ASUS, Dell, HP, Lenovo (Lenovo), Microsoft Surface and MSI, with models from Acer and GIGABYTE to follow. Trade coverage cited more than 30 laptop models and over 10 desktop systems planned for the initial wave. NVIDIA described very thin and light laptop designs, as slim as 14 millimeters and as light as about three pounds, in 14-inch and 16-inch sizes with color-accurate tandem OLED displays and G-SYNC.[1][2][4]
NVIDIA did not disclose pricing at announcement, and final prices are set by the individual OEMs.
At the keynote NVIDIA presented RTX Spark not as a one-off product but as the first generation of a permanent PC product family, with a desktop, laptop and workstation planned for each future architecture generation. The company showed a multi-generation roadmap on stage that extends beyond the current Grace Blackwell parts.[8][9]
The following roadmap reflects forward-looking statements and dates that NVIDIA presented as plans; the future generations are not shipping products, and details are subject to change.
| Generation | CPU + GPU architecture | Memory | Reported timeframe |
|---|---|---|---|
| RTX Spark (current) | Grace + Blackwell | LPDDR5X | 2026 |
| Vera Rubin Spark | Vera + Rubin (Vera Rubin) | LPDDR6 | 2027 to 2028 |
| Rosa Feynman Spark | Rosa + Feynman | next-generation memory | 2029 to 2030 |
NVIDIA's roadmap pairs each future client Spark with the corresponding generation of its broader platform, so the consumer parts track the same Vera Rubin and Rosa Feynman names NVIDIA uses for its data-center and workstation systems, including the move to LPDDR6 memory in the Vera Rubin generation. Press outlets reported slightly different exact years for the later parts because NVIDIA's slide listed them as windows rather than single dates, which is why the table above gives ranges.[8][9][10]
RTX Spark marks NVIDIA's most direct entry into the mainstream Windows PC processor market, a space long dominated by AMD and Intel x86 chips and, more recently, by Arm-based parts from Qualcomm. By combining a custom Grace Arm CPU with its own Blackwell GPU and a unified-memory design, NVIDIA is applying the architectural approach behind its data-center success to consumer hardware, and pairing it with Microsoft's Windows roadmap to push a vision of the PC as a local agentic AI platform. Commentators also compared the unified CPU-GPU-memory design to Apple silicon, while noting that NVIDIA's stated multi-generation commitment is meant to signal that the effort is durable rather than a single experiment.[1][4][8]