At its keynote address at GTC Taipei 2026, NVIDIA officially unveiled its new super chip, "NVIDIA RTX Spark," which features a custom design by MediaTek and deep collaboration with Microsoft. It integrates NVIDIA Blackwell architecture GPUs and a highly energy-efficient custom Grace CPU, and boasts up to 1 Petaflop of AI computing performance and 128GB of unified memory.
This super chip boasts that it will completely overturn the traditional Windows PC click application operation logic. With the comprehensive reconstruction of Microsoft's underlying security mechanism, it will allow the traditional PC to officially evolve from a simple "tool" into a "teammate" that can securely execute various proxy services locally.
RTX Spark: Developed in collaboration with MediaTek
RTX Spark, previously codenamed "N1X," was developed in collaboration with NVIDIA, MediaTek, and Microsoft, and designed for the Windows operating system environment. It integrates a Blackwell GPU with 6144 CUDA cores and features fifth-generation Tensor Cores supporting FP4 computing precision. The custom NVIDIA Grace CPU with up to 20 cores was custom-designed by MediaTek and "connected" to the Blackwell GPU via NVLink-C2C interconnect technology.
Overall, RTX Spark can deliver up to 1 Petaflop of AI computing performance and is equipped with up to 128GB of unified memory, enabling thin and light laptops and mini desktops to directly handle large language models with up to 120 billion (1200B) parameters on the local machine, and supporting contextual understanding with up to 100 million tokens.
Deep Integration with Microsoft: Windows Native AI Agent and OpenShell Privacy Defense
In the past, open-source personal agent projects such as OpenClaw and Hermes Agent, which became popular in the GitHub community, could handle multi-step automated tasks, but they were difficult to popularize because they could not ensure privacy and data security on a person's main PC.
To this end, NVIDIA has entered into a deep collaboration with Microsoft:
• Windows native security isolation:It introduces new Windows security and container primitives, providing identity verification and end-to-end security, allowing AI agents to operate natively in a fully controlled security sandbox.
• NVIDIA OpenShell Runtime:This is the software core of the aforementioned security architecture. Users can define the permission boundaries of the agent through OpenShell (defining what can and cannot be done), enabling it to intelligently determine privacy permissions and automatically route sensitive questions to the local model for processing. If it is necessary to send the data to a large model in the cloud, OpenShell can also automatically obfuscate and disguise personal data, eliminating users' privacy concerns.
• Native integration of work columns:NVIDIA and Microsoft's collaboration will be further expanded in the future, allowing these secure Windows AI agent experiences to be directly integrated into the Windows Taskbar user interface for easy intelligent scheduling across applications.
The software ecosystem has been completely overhauled: Adobe's core restructuring and open-source model have been accelerated twofold.
With the massive memory bandwidth of RTX Spark and the computing power of Blackwell GPUs, professional software and the open-source AI ecosystem have also undergone a comprehensive performance overhaul this fall:
• Adobe software underlying reconstruction:Adobe has announced a redesign of Photoshop and Premiere Pro for RTX Spark. The Firefly-powered Generative Fill and Generative Extend tools will see a 2x speedup on RTX Spark, while Premiere Pro will feature a completely new video editing workflow that will fully utilize the RTX Spark's 128GB unified memory for real-time editing and color grading of 12K 4:2:2 video.
• The throughput of llama.cpp has doubled:NVIDIA collaborated with the open-source community to incorporate multi-token prediction speculative decoding technology into llama.cpp. When running a large 27B language model in Qwen 3.6, local inference throughput doubled (and increased by 1.6 times for a 35B model). It also supports tensor parallelism technology for dual-GPU users.
• H Company Computer Operation Tool Acceleration:In collaboration with H Company, the AI agent optimized its Holo Computer Use model and toolkit. Even in traditional software without API support, the AI agent can automate tasks by simply "looking at the screen and operating the mouse and keyboard," just like a human. After NVIDIA quantization and optimization, execution speed doubled, and memory usage was reduced by a significant 35%.
• Blender integration with DLSS 4.5:The Blender 5.3 renderer, expected to launch this fall, will directly integrate NVIDIA DLSS 4.5 Ray Reconstruction technology, allowing 3D artists to instantly see near-final render quality physical images as they rotate the viewpoint within the window.
• ComfyUI's frame interpolation technology:Introducing the new RTX Video Frame Generation technology, which can double or even quadruple the frame rate of low-frame-rate videos (typically 15-20 fps) generated by AI diffusion models on the local device, making AI videos instantly smooth.
This fall's hardware lineup is set to break the heavy burden of performance PCs.
Thanks to the chipset's excellent power consumption control, RTX Spark will break the traditional fate of high-performance PCs having to maintain a bulky hardware design.
The laptop's body thickness is kept within 1.4 cm and its weight is only about 1.36 kg. It has all-day battery life and a Tandem OLED (dual-layer tandem OLED) screen with G-SYNC technology.
Starting this fall, ASUS will launch its first wave of RTX Spark application products under the ProArt brand, Dell will launch the XPS 16 Creator Edition, HP will launch the OmniBook series, and Microsoft will launch the Surface Laptop Ultra equipped with RTX Spark. In addition, Lenovo and MSI will also launch the first thin and light laptops and ultra-efficient small desktops equipped with RTX Spark. Acer and Gigabyte will also launch corresponding models later.
In addition, NVIDIA also announced the simultaneous launch of DGX Station for Windows, equipped with data center-level GPUs, for enterprise and Linux development environments, bringing the computing power of AI supercomputers directly to developers' desks.
Analysis: NVIDIA's partnership with MediaTek and Microsoft poses a challenge to the traditional "Wintel" system.
Judging from the RTX Spark strategy developed by NVIDIA in collaboration with Microsoft and MediaTek, this is undoubtedly a challenge to the traditional PC and chip market.
Over the past two years, the Windows on Arm camp, led by Qualcomm, has made a name for itself in the Copilot+ PC market with its long battery life. However, due to the lack of powerful hardware GPU computing power, it has struggled to handle local models with billions of parameters, professional 3D rendering, and AAA games. This time, NVIDIA has paired the Grace CPU with the Blackwell GPU, along with up to 128GB of unified memory, not only solving the problems of "mobile battery life" and "productivity hardware bottlenecks," but also further enhancing "local AI privacy and security."
When your laptop doesn't need to be constantly connected to the internet or pay cloud subscription fees, and can run Adobe generative tools, run 100 billion parameter models, and even play AAA games locally at twice the speed, it means that the "click application, click mouse" behavior logic of the PC industry for the past 40 years has officially come to an end.
Microsoft's willingness to relinquish underlying security isolation permissions for NVIDIA, and even integrate OpenShell into its worklist, demonstrates that in Microsoft's view, the complete deployment capability of local AI has surpassed the old norm of alliances with Intel, AMD (and even Qualcomm). The launch of this "Personal AI PC" will undoubtedly drop a huge bombshell on the entire PC and processor market this fall.
With the launch of the RTX Spark super chip, NVIDIA has announced the official addition of a new member to its product line, which is expected to bring greater market development opportunities.
