NVIDIA has unveiled the details of its new GB10 Superchip, a compact powerhouse designed to bring datacenter-level AI capabilities into desktop-sized machines. 
First shown inside the DGX Spark, the company’s AI PC concept, the GB10 combines cutting-edge CPU and GPU technology with an efficient multi-die package that fits neatly into a workstation form factor.
The chip is built on TSMC’s 3nm process and is split into two dielets: the S-dielet, which carries 20 ARM v9.2 CPU cores (arranged in two clusters of 10 with 32 MB total L3 cache), and the G-dielet, home to the GB100 Blackwell GPU. The GPU is integrated as an iGPU but doesn’t compromise on performance, boasting fifth-gen Tensor Cores with DLSS 4, RTX Ray Tracing, 31 TFLOPs FP32, and an astonishing 1000 TOPS of NVFP4 compute for AI workloads.
Memory is another standout. The GB10 supports up to 128 GB of LPDDR5x-9400 unified system memory, offering 301 GB/s bandwidth and system-wide coherency through CHI-E protocol. The GPU and CPU share up to 600 GB/s aggregate bandwidth via NVIDIA’s C2C interface, while a 16 MB system-level cache boosts data efficiency. The result: workloads ranging from fine-tuning 70B parameter models to handling 200B parameter giants are possible in a single compact box.
The DGX Spark workstation doesn’t stop at raw compute. With ConnectX-7 networking, two units can link together to work with models surpassing 400B parameters, blurring the line between a desktop and a personal cloud cluster. The platform runs DGX Base OS and NVIDIA’s AI software stack, allowing seamless transitions between local and cloud deployments. Despite its 140W TDP, the system supports multi-head displays, PCIe and USB expansion, and up to 8K/120Hz output over HDMI 2.1a.
NVIDIA highlights its partnership with Mediatek, whose CPU IP underpins the system. Extensive modeling of GPU memory traffic with Mediatek’s subsystem made the integration possible. While DGX Spark targets developers and professionals, the GB10 Superchip is widely expected to trickle down into consumer laptops and Mini PCs, giving us a glimpse of what next-gen AI PCs will look like.
3 comments
lol slow af, even zen cores will wreck this thing 😂
bro quit hypin, amd just posted 2 quarters loss in datacenter lol nobody want their ai gpus
Waste of sand tbh, just another overpriced toy 🤷♂️