Hardware overview
For years, PC experiences centered on opening applications and completing tasks through clicks, menus, and keystrokes. As AI becomes more integrated into everyday workflows, the interaction model is shifting toward natural-language prompts, AI-assisted task completion, and local processing that can help keep work responsive and efficient. NVIDIA RTX Spark™ is a silicon platform designed with this shift in mind, helping Windows PCs support experiences such as local AI agents, advanced models, and creative workloads.
The following components work together to help power these AI-enabled experiences.
Blackwell RTX GPU
RTX Spark™ includes an NVIDIA Blackwell RTX GPU with 6,144 CUDA (Compute Unified Device Architecture) cores. CUDA cores are small processing units that help accelerate graphics and parallel computing tasks. This GPU capacity helps support demanding workloads such as 3D rendering, AI model acceleration, high-resolution media processing.
For workflows that rely on intensive visual processing—such as multi-stream 8K video editing, 3D content creation, or local image and video generation—this graphics capability helps bring desktop-class performance into a more portable form factor.
NVIDIA Grace CPU
RTX Spark™ features an NVIDIA Grace CPU (Central Processing Unit) with 20 cores. The CPU is designed to manage general system activity and distribute workloads efficiently.
This design helps the system stay responsive while handling background tasks, code compilation, browser activity, and other processes alongside GPU-intensive workloads. By balancing performance and power use, the CPU also helps manage thermal output and support battery efficiency.
Up to 128 GB unified memory
RTX Spark™ supports up to 128 GB of unified memory. Unified memory allows the CPU and GPU to access a large, shared pool of high-speed memory, rather than relying on separate system and dedicated graphics memory.
This memory architecture enables large local AI models, complex 3D scenes, and data-intensive creative workloads by dynamically sharing memory between the CPU and GPU.
Up to 1 Petaflop FP4 AI performance
RTX Spark™ delivers up to 1 petaflop of FP4 AI performance. A petaflop represents one quadrillion operations per second, and FP4 refers to a compact 4-bit floating-point format designed to help AI models run more efficiently. This performance enables responsive local experiences such as AI-assisted code generation, document summarization, data retrieval, and voice-based interactions.
Why this matters: RTX Spark™ can support more demanding workloads directly on the device. For organizations, this can mean faster access to AI-assisted experiences, improved responsiveness, and more flexibility to run AI tasks locally when performance, connectivity, or considerations are important.