At Hot Chips 2025, NVIDIA offered a detailed look at the GB10 "Grace Blackwell" superchip, presenting a compact multi-die design meant to bring datacenter capabilities to a desktop-sized workstation. The chip pairs a MediaTek-sourced Arm CPU die with a Blackwell GPU die in a 2.5D package built on TSMC's 3 nm process. The CPU side implements 20 Arm v9.2 cores split into two clusters of ten, each cluster backed by a 16 MB shared L3 (32 MB total) while each core retains private L2 storage. The memory subsystem is a unified LPDDR5X-9400 fabric on a 256-bit bus, supporting up to 128 GB and delivering roughly 301 GB/s of raw bandwidth to the package. High-speed I/O is concentrated on the CPU die, while NVMe storage and peripherals ride PCIe lanes there, and a ConnectX-7 NIC is attached via a PCIe Gen 5 x8 link to enable multi-unit networking.

NVIDIA described the GPU die as a scaled Blackwell configuration tuned for low-power, small-form-factor operation. The quoted peak throughput is approximately 31 TeraFLOPS for FP32 and around 1,000 TOPS when using NVIDIA's NVFP4 reduced-precision format. The GPU carries a large 24 MB L2 that can act as a higher-level cache visible to the CPU, creating a coherent cache hierarchy across the two dies. NVIDIA states that the inter-die C2C link offers aggregate bandwidth on the order of 600 GB/s, enabling low-latency sharing without requiring heavy software-managed copying. The package is rated at approximately 140 W TDP and exposes multi-display outputs (DisplayPort alt-mode plus HDMI 2.1a options) alongside security and virtualization features intended for professional workloads.