NVIDIA Unveils GB10 Superchip Details for DGX AI PCs: 3nm Technology, 20 ARM v9.2 CPU Cores, 1000 TOPS NVFP4 Blackwell GPU, LPDDR5x-9400 Memory Support, and 140W TDP

NVIDIA has recently unveiled its state-of-the-art GB10 Superchip, which utilizes the Blackwell GPU architecture, powering the latest DGX AI Mini supercomputers.

NVIDIA GB10 Superchip: A Revolutionary SoC and GPU Architecture

The spotlight is currently on NVIDIA’s DGX Spark, the pioneering system equipped with the GB10 Superchip, marking a significant entry into the “AI PC”market. Following its announcement, numerous companies have begun formulating their own AI PC platforms leveraging the GB10 technology. During the Hot Chips 2025 conference, NVIDIA offered an in-depth exploration of the GB10 Superchip, demonstrating how it effectively scales the Blackwell architecture for Mini developer stations and workstations.

Diagram of NVIDIA Blackwell GPU architecture with data center integration.

The DGX Spark was conceptualized to serve as a Mini AI Supercomputer, built around the advanced Blackwell architecture.

Diagram of building a mini AI supercomputer with Blackwell technology features and DGX Spark.

Key Features of the DGX Spark Workstation

The DGX Spark Workstation incorporates several standout features:

GB10 Grace Blackwell Superchip: Optimized for AI, Data Science, Compute, Rendering, and Visualization tasks.
128GB Coherent Unified System Memory: Capable of handling large AI models with up to 200 billion parameters and fine-tuning models with up to 70 billion parameters.
ConnectX-7 Networking: Facilitates interconnection between two DGX Spark systems, allowing collaboration on models up to 405 billion parameters.
DGX Base OS and NVIDIA AI Software Stack: This ensures smooth transfer of workloads between DGX Spark and DGX Cloud or any accelerated cloud infrastructure.
Versatile Deployment Options: Can be configured as either an AI Workstation or a network-connected personal AI cloud.
Enhanced User Experience: Offers support for multi-head displays and various connectivity options.
Compact and Energy-Efficient Design: Fits comfortably on any desk and operates with a standard wall outlet.

NVIDIA DGX Spark Workstation features powerful AI capabilities and compact, efficient design.

Diving Deeper: Specifications of the GB10 Superchip

Examining the GB10 Superchip further, we find that its System-on-Chip (SoC) is engineered with two dielets: the S-Dielet, which houses the CPU and memory subsystem, and the G-Dielet, which is dedicated to the GPU core. Both dielets utilize Advanced 2.5D packaging technology and are manufactured using TSMC’s cutting-edge 3nm process.

NVIDIA GB10 Superchip: AI performance specs including 20 Arm Cores and 128GB memory.

The CPU architecture is based on ARM’s v9.2 design, featuring a total of 20 cores, configured into two clusters. Each core benefits from a dedicated L2 cache alongside a collective L3 cache of 32 MB shared between the clusters.

NVIDIA GB10 Specifications Sheet with CPU, GPU, and memory details.

The GPU portion of the chip, leveraging the GB100 Blackwell architecture, functions as an integrated GPU (iGPU) due to its placement within the same chip package. It boasts advanced features including 5th Generation Tensor Cores with DLSS 4 support and RTX Ray Tracing cores, delivering an impressive computational capability of up to 31 TFLOPs for FP32 and 1000 TOPS for NVFP4 (FP4) operations tailored for AI applications. Additionally, the GPU includes a 24 MB L2 cache to enhance performance.

In terms of memory capabilities, the NVIDIA GB10 Superchip supports 256b LPDDR5x (Unified Memory Architecture) with access speeds reaching 9400 MT/s, translating to a remarkable raw bandwidth of up to 301 GB/s. This setup enables a maximum memory capacity of 128 GB. Its high-performance coherent fabric is tuned for CHI-E Coherency Protocol, supplying the GPU with an extensive aggregate system bandwidth of 600 GB/s over the C2X interface.

NVIDIA GB10 specs: Display, security, TDP, and OS details with chip graphic.

Additionally, the GB10 features a 16 MB System Level Cache, designed as L4 for the CPU, which promotes power-efficient data-sharing across multiple SoC engines. The high-bandwidth, low-power C2C interface leverages NVIDIA’s NVLINK technology.

Connectivity options are robust, with support for PCIe, USB, and Ethernet over PCIe. It can handle up to four simultaneous displays (3 DisplayPort + 1 HDMI) at resolutions of up to 4K at 120Hz using DP Alt-mode, and even supports 8K at 120Hz with HDMI 2.1a. Security is prioritized as well, featuring Dual Secure Root support, an SROOT processor, OSROOT processor, and compatibility with both fTPM and discrete TPM. The chip operates with a total design power (TDP) of 140W.

Diagram of GB10 scalability with NVIDIA ConnectX for multi-GPU computing efficiency.

Scalability of the GB10 Superchip

Another noteworthy feature of the GB10 Superchip is its scalability. Multiple GB10 chips can be interconnected via NVIDIA’s ConnectX Technology, enabling enhanced throughput, bandwidth, and DRAM capacities crucial for supporting larger AI models. Each ConnectX Network Interface Card (NIC) connects to the GB10 SoC through a PCIe Gen5 x8 interface, with units communicating over an Ethernet framework.

Collaboration diagram between NVIDIA and Mediatek, merging CPU and GPU architectures for efficiency.

NVIDIA describes the GB10 Superchip SoC as a successful partnership with Mediatek, utilizing Mediatek’s CPU IP. The chip underwent detailed performance modeling of GPU memory interactions within Mediatek’s memory infrastructure.

NVIDIA DGX Spark computer with display, optimizing AI and GPU performance for developers.

Future Prospects of the GB10 Superchip

The most exciting aspect of the GB10 Superchip is its potential to eventually reach consumer markets, such as laptops and Mini PCs. Reports suggest that the upcoming N1X and N1 SoCs may serve as the first consumer-focused NVIDIA SoCs, with the GB10 providing a glimpse into the capabilities these chips could offer.

Source & Images