NVIDIA RTX Spark Architecture Enables Local Artificial Intelligence and High Performance Gaming on Thin Laptops with Secure Private Execution Environments
NVIDIA has revealed its new superchip architecture: the RTX Spark. This new silicon is set to transition Windows PCs from traditional applications hosts to intelligent agents of local artificial intelligence. The chip is the sum of over thirty years of computation research from the industry leading company, blending known standards like CUDA, G SYNC and TensorRT into a single package. Aimed at thin laptops and compact desktop computers, the chip offers sustainable, high end performance for personal artificial intelligence agents.
The design of the silicon is an NVIDIA Blackwell RTX GPU containing 6144 CUDA cores and 5th generation Tensor cores with FP4 precision. The graphics system is interconnected with a NVLink C2C chip to chip interconnect to a 20 core Grace CPU. To achieve this remarkable energy efficiency the company worked closely with MediaTek to leverage their Arm system on a chip know how and to optimize the system. The integrated hardware provides an amazing 1 petaflop local intelligence processing and 128 GB of unified memory which is crucial for processing and manipulating large local models.
Founder and CEO Jensen Huang described the architectural shift:
"The PC is being reinvented. For 40 years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask, and the PC does the work. RTX Spark brings everything NVIDIA has built, CUDA, RTX, our AI platform, into a single superchip. Local agents. Frontier models. Creative workflows. RTX games. All on a laptop. This is the new PC. The personal AI computer."
For all its promise, one of the biggest issues for wide adoption is privacy and data protection. In most existing cloud based systems, a significant volume of user data must be sent over the internet to external servers, something which NVIDIA and Microsoft have tackled in tandem with a secure execution environment inside Windows. This security framework will utilize new Windows primitives and the NVIDIA OpenShell runtime, which allow the user to define exactly what an agent is able to do and access. It can either run entirely local and open source AI agents such as the OpenClaw or Hermes Agent, or partially mask sensitive personal information prior to sending more complex requests to the cloud.
Vincent Koc, the Chief Architect of the OpenClaw Foundation, spoke of the importance of a unified security system for local agents:
"We are strong supporters of deploying agents like OpenClaw securely into the Windows ecosystem. Running solutions like OpenShell and the Microsoft security primitives on RTX Spark will enable users to leverage a fully integrated stack for private, personal agents running on device."
And Microsoft's Chairman and CEO, Satya Nadella, enthusiastically supported the effort, stating it marks a significant milestone towards bringing unmetered, private intelligence into ordinary office environments.
Not content with just running complex local models, the powerful and integrated memory and compute capabilities of the RTX Spark also allow for extremely demanding creative workloads as well as gaming. The superchip is capable of processing 90 GB virtual scenes using the OptiX ray tracing engine, of decoding 12 K resolution video files, and of running massive language models consisting of 120,000,000,000 parameters with a 1 million token context window. When utilized for games the hardware, at 1440p resolution, can manage over 100 fps while using G SYNC and ray tracing technologies.
Software developers are already getting their tools on board; Adobe is developing a new version of Premiere and Photoshop to work with the integrated memory structure, and this is expected to result in a 2x performance improvement for tools like Firefly. Other tools such as ComfyUI and OTOY Octane are being adapted to take advantage of the huge local memory capacity allowing users to run large diffusion models and neural rendering tools locally on a laptop without needing to stream rendering from cloud based networks.
Devices will be a maximum of 14 millimeters thin and will weigh a minimal three pounds. The thin laptops and compact desktops will be available with 14 to 16 inch screen sizes, complete with high quality, color accurate OLED screens. Initial devices powered by this new superchip are expected to arrive in autumn, and are being developed by ASUS, Dell, HP, Lenovo, MSI, and Microsoft Surface.
