NVIDIA Blackwell Solutions

Two Dies in a Single GPU

NVIDIA Blackwell GPU is one unified chip built from two Blackwell GPU dies. By merging two dies over a 10TB/s chip-to-chip interface on a single substrate, NVIDIA developed the largest and most performant GPU today.

NVIDIA Blackwell architectural innovation translates to the astronomical throughput with a single GPU capable of 20 petaFLOPs in FP4 and 5 petaFLOPs in FP16, the same throughput as an NVIDIA DGX A100 system.

• 2nd Gen Transformer Engine: Blackwell Tensor Cores utilizes fine-grain scaling techniques called micro-tensor scaling to optimize performance and accuracy, enabling 4-bit floating point (FP4).
• NVLink & NVSwitch: 5th NVIDIA NVLink interconnect can scale up to 576 GPUs, and the NVIDIA NVLink Switch Chip further scales to multi-server clusters at up to 1.8TB/s interconnect.