NVIDIA Launches DGX Spark Micro-Superchip GB10

Recently, NVIDIA unveiled its latest micro-superchip, the GB10, at Hot Chips 2025, bringing data-center-class performance to developers for local AI applications.

The GB10 combines a MediaTek CPU with an NVIDIA GPU, manufactured on TSMC's 3nm process and integrated using 2.5D packaging. NVIDIA's NVLink technology enables a staggering 600GB/s bidirectional bandwidth between chips, ensuring lightning-fast communication.

On the CPU side, GB10 packs 20 ARM v9.2 cores split evenly between X925 and Cortex A725 clusters, along with 32MB L3 and 16MB L4 caches to optimize GPU interaction. While full GPU specs remain under wraps, peak performance is expected to reach 1 PetaFLOP FP4 or roughly 31 TFLOPS in FP32 precision.

Performance-wise, the GB10 rivals the RTX 5070 but consumes only 140W compared to the RTX's 250W. Its 128GB memory far surpasses the RTX 5070's 12GB, giving it a clear edge for handling large AI models. Instead of high-bandwidth memory (HBM), the GB10 uses LPDDR5x running at 9400MT/s on a 256-bit memory bus, delivering 273–301 GB/s of bandwidth.

The DGX Spark platform supports both edge inference and AI model prototyping. NVIDIA highlights that 128GB of memory can fine-tune models with up to 70 billion parameters and handle inference for models with up to 200 billion parameters. For even larger workloads, dual 200GbE ports via ConnectX-7 allow tasks to span two DGX Spark units, scaling inference and fine-tuning seamlessly.

Because the GB10 shares the same technology foundation as NVIDIA's larger GB200 and GB300 chips used in data centers, developers can build prototypes on DGX Spark and later deploy them to large-scale environments without rewriting code—streamlining the path from development to production.