Particle.news

Nvidia Unveils RTX Spark Superchip for AI-Driven Windows PCs

The chip is built to run large language models and local AI agents on thin, battery-powered laptops to move key workloads off cloud servers and onto users' devices.

Overview

  • Nvidia announced RTX Spark at its Computex/GTC keynote on Monday in Taipei as a full-system 'superchip' for Windows laptops and compact desktops designed to run advanced AI models and on-device agents.
  • The platform pairs a Blackwell GPU with 6,144 CUDA cores, a 20-core Grace-based CPU, up to 128 GB of unified LPDDR5X memory and about 1 petaflop of AI throughput, which Nvidia says can run models up to roughly 120 billion parameters locally.
  • Nvidia named MediaTek as a CPU design partner and said Microsoft helped optimize Windows runtime and security for the platform, with OEMs including Microsoft Surface, Dell, HP, Lenovo, ASUS, MSI, Acer and Gigabyte planning Spark systems for autumn release.
  • The company said its Vera CPU is entering production and is already being trialed by AI firms, and it also unveiled a Unitree research robotics platform that combines Unitree's H2 Plus robot with Nvidia Jetson Thor hardware and Isaac GR00T tools.
  • Investors pushed Nvidia shares up about 6% after the announcements while rival chip stocks fell, and analysts flagged open questions about pricing, supply constraints, real-world agent usefulness and software and privacy controls for on-device AI.