How NVIDIA’s Blackwell GB200 NVL72 Redefines AI Compute with 10 TB/s Interconnect

Overview of the Blackwell GB200 NVL72 Platform

NVIDIA’s latest Blackwell platform introduces the GB200 NVL72 GPU, built on a 4‑nm process with 208 billion transistors and a dual‑die design that delivers up to 10 TB/s of full‑mesh inter‑GPU bandwidth via the fifth‑generation NVLink technology.

Performance Gains

Compared with the previous Hopper generation, the HGX B200 system achieves a three‑fold increase in training throughput and a 15‑fold boost in inference performance for large models such as GPT‑MoE‑1.8T. The second‑generation Transformer engine, FP8 precision, and optimized Tensor Core architecture together provide a three‑times faster training speed for the same model.

NVLink‑C2C Interconnect

The NVLink‑C2C link offers 1.8 TB/s per GPU‑to‑GPU connection, 130 TB/s aggregate bandwidth in a full‑rack configuration, and double the bandwidth of PCIe Gen5. This high‑speed, low‑latency fabric enables up to 18 NVLink 100 GB/s links per GPU, effectively scaling multi‑GPU clusters for trillion‑parameter AI workloads.

DGX SuperPOD and Rack‑Level Design

The DGX SuperPOD integrates 36 × GB200 Grace‑Blackwell Superchips per rack, each combining a Grace CPU with two Blackwell GPUs. Liquid‑cooled chassis reduce power consumption while delivering up to 80 PFLOPS of AI performance and 1.7 TB/s of high‑bandwidth memory. The system leverages NVIDIA Quantum‑InfiniBand and NVLink to interconnect multiple racks, supporting exaFLOP‑scale AI training and real‑time inference.

Copper‑Cable Interconnect Trend

NVL72 adopts a copper‑direct‑connect solution that offers lower cost and simpler routing compared with optical alternatives. Industry analysis predicts copper‑backplane connectors will become mainstream for AI servers, with bandwidths scaling from 10 Gbps to 224 Gbps in the near future.

Market Implications

The combination of massive performance uplift, efficient cooling, and cost‑effective copper interconnect positions the Blackwell platform as a key driver for next‑generation AI workloads, large‑scale simulations, and data‑center analytics. Domestic suppliers such as 华丰科技 and 中航光电 are gaining market share by providing high‑speed backplane solutions, challenging traditional overseas vendors.