Overview

• GB10 is a 3nm, dual‑dielet SoC for DGX Spark and AI PC platforms with a 20‑core Arm v9.2 CPU, a Blackwell iGPU, and up to 128GB coherent LPDDR5X UMA in a 140W package.
• The integrated Blackwell graphics block delivers up to 31 TFLOPs FP32 and 1,000 TOPS NVFP4, with a chip‑to‑chip NVLink‑based interface exposing up to 600 GB/s aggregate bandwidth to the GPU.
• NVIDIA says GB10’s S‑die uses MediaTek CPU IP, with the collaboration relying on extensive co‑simulation and a custom NVLink C2C link to unify the CPU, memory subsystem, and GPU.
• GB300 is presented as a dual‑die accelerator with 208 billion transistors, 20,480 CUDA cores, NV‑HBI interconnect rated at 10 TB/s between dies, and up to 288GB HBM3E delivering 8 TB/s per GPU, with NVIDIA citing roughly 50% faster AI performance versus the prior generation.
• Rack‑scale GB300 NVL72 systems are described at up to 1.1 exaFLOPS dense FP4, and CoreWeave reports a DeepSeek R1 run needing 4 GB300s versus 16 H100s with about 6× higher raw throughput per GB300 GPU.