NVIDIA DGX Spark is equipped with the GB10 chip co-designed with MediaTek and will be launched on 10/15

NVIDIA (NVIDIA)’s upcoming NVIDIA DGX Spark personal AI supercomputer is equipped with the GB10 Grace Blackwell super chip designed by MediaTek and NVIDIA. NVIDIA DGX Spark allows developers to prototype, fine-tune and infer large-scale AI models locally. DGX Spark will be launched on October 15 and is expected to drive a new wave of AI development in various industries.

MediaTek pointed out that the Grace CPU in the GB10 Grace Blackwell super chip uses a 20-core Arm architecture, which fully demonstrates MediaTek’s technical strength in high-performance, low-power design, memory subsystems and high-speed interfaces. Combining the latest generation Blackwell GPU with 128GB of unified architecture memory, GB10 provides up to 1 PFLOP of AI computing power to significantly accelerate model tuning and instant inference.

In addition, this configuration allows developers to run large AI models with up to 200B parameters, or use the built-in ConnectX-7 network interface card to connect two DGX Spark systems in series to run model inference with up to 405B parameters. DGX Spark packs powerful performance into a compact body that fits easily on your desk, while its low-power design requires only a standard outlet to operate.

Vince Hu, deputy general manager of MediaTek Data Center and Computing Business Group, said that DGX Spark will lead AI prototyping into a new era, further realizing our mission of "making technology more accessible from the edge to the cloud" while effectively solving performance and power consumption challenges. The GB10 super chip combines MediaTek's efficient computing design in the data center field and energy-saving technology for consumer devices, and is specifically designed for AI workloads.

The GB10 cooperation is based on the deep cooperation between MediaTek and NVIDIA in multiple fields, bringing advanced AI technology to markets such as ultra-large-scale data centers, Internet of Things applications, and software-defined cars.