NVIDIA RTX Spark Superchip is the first NVIDIA processor that aims to change PCs forever
After Grace and Vera dedicated to the data center world and the GB10 Superchip used in DGX Spark systems for AI developers, NVIDIA officially enters the PC processor market with the RTX Spark Superchip, a new Arm platform intended for compact notebooks and desktops that will debut in the fall.
Unveiled at Computex 2026, RTX Spark represents NVIDIA's most ambitious attempt to redefine the very concept of personal computers, uniting in a single chip Arm CPU, Blackwell GPU, and AI acceleration to create systems designed not only for gaming and productivity but especially for local execution of AI agents.
A chip with 70 billion transistors featuring Arm CPU and Blackwell GPU
NVIDIA has not yet revealed the full range of RTX Spark products but has outlined the features of the top configuration, derived directly from the GB10 Superchip already used in DGX Spark systems.
The SoC integrates a 20-core Arm CPU based on Grace architecture and a Blackwell GPU with 48 Streaming Multiprocessors, equivalent to 6,144 CUDA cores and fifth-generation Tensor Cores with support for FP4 precision. The chip is manufactured using TSMC's 3-nanometer process and contains approximately 70 billion transistors in total.
NVIDIA confirmed its collaboration with MediaTek, which contributed to the development of the Arm component, with particular attention to energy efficiency, connectivity, and power consumption optimization.
According to the company, the graphical performance of the flagship solution is comparable to that of a GeForce RTX 5070 Laptop GPU, but with a significantly lower power consumption due to the project's high level of integration.
Up to 128 GB of unified memory and NVLink-C2C
One of the most interesting aspects of the new platform concerns the memory subsystem. RTX Spark can be equipped with up to 128 GB of unified memory and a bandwidth of 300 GB/s. NVIDIA claims this is the largest amount of memory directly addressable by the GPU ever seen on a consumer RTX system.
The CPU and GPU communicate through NVLink-C2C, the chip-to-chip interface already used in the company's professional AI products. The connection offers a bandwidth of 600 GB/s, about five times greater than that available with PCI Express 5.0, while maintaining low power consumption.
This configuration allows the platform to locally execute language models with up to 120 billion parameters with context windows of up to a million tokens, eliminating many of the limitations imposed by traditional Windows notebooks.
The RTX Spark platform also integrates a dedicated NPU to meet the requirements of Microsoft’s Copilot+ PC, although the main AI acceleration continues to rely on the Tensor Cores of the Blackwell GPU.
NVIDIA and Microsoft aim to transform the PC into an AI agent
More than hardware specifications, much of the presentation focused on the concept of a "personal AI computer". According to NVIDIA, the future of PCs will no longer be a passive tool that runs applications launched by the user, but a system capable of operating as a true AI agent that can understand requests, interact with different software, and perform tasks autonomously.
Therefore, NVIDIA is closely collaborating with Microsoft to create a Windows platform specifically designed for local AI agents. The project includes new integrated security primitives in Windows and the introduction of NVIDIA OpenShell, a runtime developed to enable the secure execution of agents directly on the user's device.
According to both companies, OpenShell will precisely define what agents can or cannot do, route requests to local or cloud models based on privacy preferences, and even mask sensitive information before sending it to external services. The goal is to create an ecosystem where AI agents can operate within Windows, giving the user full control.
NVIDIA has identified initial partners such as OpenClaw and Hermes Agent, which are developing applications capable of automating workflows between Windows applications, searching for files in natural language, generating images and videos, or writing code directly on the local system.
Microsoft also plans to integrate new AI agent-related features directly into the Windows interface, making them accessible from the taskbar.
Gaming and content creation without compromise
On the creative front, NVIDIA claims that RTX Spark can handle 3D scenes larger than 90 GB using OptiX and DLSS, edit 12K 4:2:2 video using the Blackwell encoder and decoder, and generate AI content directly on the local machine.
The company also promises enough performance to play AAA titles at 1440p and over 100 FPS with ray tracing, DLSS, and Reflex enabled. The platform will also support future evolutions of the RTX ecosystem, including DLSS 4.5 Ray Reconstruction based on second-generation transformer models and RTX Video with Frame Generation 4x.
According to NVIDIA, over 100 software houses are already working on supporting the new platform. These include Adobe, Blackmagic Design, Blender, CapCut, ComfyUI, OTOY, and various game developers such as Riot Games, Remedy Entertainment, NetEase, KRAFTON, and Xbox.
On the gaming front, NVIDIA has assured that it is working "closely" with developers and anti-cheat solution providers to ensure that all the most popular titles run perfectly. "For non-native Arm apps and games, our multi-year collaboration with Microsoft has allowed us to further enhance the Prism emulator, already highly efficient, by adding advanced features such as AVX2 and new performance optimizations. ISVs are excited about Spark and are doing everything they can to support it," said Mark Averman, Consumer Product Marketing Lead for RTX Spark.
Adobe redesigns Premiere and Photoshop for RTX Spark
One of the most significant announcements regarding collaboration with Adobe. The company is redesigning core components of Premiere Pro and Photoshop from the ground up to leverage unified memory, the Blackwell GPU, and TensorRT optimally.
In the case of Premiere, a new video pipeline will be introduced to extensively utilize GPU acceleration for editing, color grading, rendering, and real-time AI features. Also, Photoshop will receive a new graphics engine optimized for GPU-accelerated compositing, with support for live filters, HDR, and new natural brush tools.
According to Adobe, the work done together with NVIDIA and Microsoft will achieve up to twice the performance in AI, editing, and graphic processing operations compared to current approaches. The two companies are also working on integrating the MCP protocol and AI agents directly within creative applications, allowing users to interact with Premiere and Photoshop via natural language.
Thin notebooks and compact desktops
The first RTX Spark systems will arrive in the fall as Windows on Arm notebooks and compact desktops. NVIDIA talks about laptops with thickness up to 14 millimeters, a minimum weight of about 1.36 kilograms, and displays ranging from 14 to 16 inches. The premium models announced at the debut include ASUS ProArt, Dell XPS, HP OmniBook, Lenovo Yoga Pro 9N, Microsoft Surface Ultra, and MSI Prestige N16.
On the desktop side, ASUS, Dell, HP, Lenovo, Acer, Gigabyte, and MSI are developing compact systems based on the new platform. All these devices will use Windows on Arm, an element that represents perhaps one of the main challenges for NVIDIA.
The company explained that it is working closely with Microsoft to improve compatibility and performance of the operating system. At the same time, NVIDIA is collaborating with AI framework developers such as PyTorch, Hugging Face, llama.cpp, and LM Studio to ensure a mature software ecosystem from the debut.
Many promises, but testing will be decisive
NVIDIA's ambitions are evident: leverage leadership in artificial intelligence and gaming to enter a market dominated for decades by Intel, AMD, and, more recently, Qualcomm. RTX Spark appears on paper as an extremely interesting proposal thanks to the combination of Blackwell GPU, high-capacity unified memory, AI acceleration, and low power consumption.
However, essential elements are still lacking to formulate a complete judgment. NVIDIA has not published detailed benchmarks nor independent performance data, while many of the features illustrated depend on the evolution of the Windows on Arm ecosystem and the optimization work by software developers. The fall market arrival leaves room to refine both the software platform and the commercial offering, in a context where the pricing trends of memory and storage could significantly impact the final cost of RTX Spark systems.