NVIDIA Details Blackwell Ultra B300: 208-Billion Transistor Dual-Die Superchip for AI Clusters

At an international engineering symposium, NVIDIA shared comprehensive hardware architectural specifications for the Blackwell Ultra B300 data center platform. Designed for trillion-parameter AI pre-training and real-time reasoning workloads, the chip incorporates twin dies interconnected across a 10 TB/s ultra-dense silicon bridge.

Key Architectural Specifications

• CoWoS Packaging: 208 billion transistors manufactured on a custom TSMC 4NP node operating as a single unified accelerator.
• HBM3e Memory: 288GB of high-bandwidth memory delivering up to 8.0 TB/s of aggregate memory throughput per accelerator socket.
• NVLink 5 Fabric: 1.8 TB/s bidirectional bandwidth per GPU, enabling clusters of up to 576 GPUs in an NVLink domain to operate as a single virtual GPU unit.
• Second-Generation Transformer Engine: Native 4-bit floating point (FP4) micro-tensor formats doubling effective inference throughput while preserving numerical stability.

The architecture sets a new compute density benchmark for hyper-scale cloud data centers worldwide.