A new desktop-scale AI system built around the co-designed GB10 Superchip bringing datacenter-grade AI power and efficiency to developers’ workstations.
NVIDIA’s newly launched DGX Spark personal AI supercomputer features the GB10 Grace Blackwell Superchip, a joint innovation with MediaTek that blends power-efficient CPU design with high-throughput GPU performance. The system is now available globally, aimed at developers and researchers seeking to prototype and fine-tune large AI models directly on the desktop.
The superchip integrates MediaTek’s expertise in high-efficiency memory subsystems and high-speed interfaces with NVIDIA’s next-generation Grace 20-core Arm CPU and Blackwell GPU architecture. The result is a compact platform delivering up to 1 petaFLOP of AI performance and 128GB of unified memory, enabling real-time inference and model training without relying on massive server clusters.
The key characteristics are:
- Supports AI models with up to 200 billion parameters on a single system.
- ConnectX-7 networking enables linking two DGX Spark units.
- Combined setup scales workloads to 405 billion parameters.
Despite its power, the system’s compact form factor and modest power requirements allow it to operate from a standard electrical outlet—making supercomputing capabilities accessible to small labs and individual developers.
Industry observers view DGX Spark as part of a broader trend toward distributed, desk-side AI computing, a shift that could reduce reliance on cloud resources and bring faster iteration to model development. The partnership also underscores the growing convergence between consumer chip efficiency and datacenter-grade AI performance.
For MediaTek, the collaboration extends a long-standing relationship with NVIDIA spanning data centers, automotive platforms, and IoT devices. The GB10 design demonstrates how MediaTek’s low-power design philosophy and system integration expertise can scale into high-performance computing. By co-developing the GB10 Superchip, both the companies are redefining what an AI workstation can achieve—placing the capability to train and test frontier-scale models directly on the developer’s desk.
