Specifications

Overview

The NVIDIA GB300 Grace Blackwell Ultra Superchip pairs two B300 Blackwell Ultra GPU dies with an NVIDIA Grace CPU on a single CoWoS-L interposer. Each superchip carries 288 GB of HBM3e at 8 TB/s — 50% more memory and bandwidth than the prior GB200 — enabling frontier-scale large-language-model training and inference with reduced interconnect overhead.

Blackwell Ultra advances the GB200 design with higher-density HBM3e, a dual-reticle GPU delivering more than 20,000 CUDA cores per die, and NVLink-C2C fabric connecting the GPU pair to the Grace CPU at 900 GB/s. It shipped in January 2026 as NVIDIA's most powerful single-superchip compute building block.

QS Compute sources GB300 Grace Blackwell Ultra Superchips and their HGX/rack configurations for enterprise and cloud AI deployments across APAC. Contact us for cluster-level pricing and availability.