At the GTC2026 Annual Conference, NVIDIA founder Jensen Huang officially unveiled the next-generation AI acceleration platform codenamed "Vera Rubin." Named after astronomer Vera Rubin, who provided evidence for dark matter, the platform marks NVIDIA's deep transformation from a single-chip supplier to a full-stack AI factory infrastructure. The Rubin GPU uses TSMC's 3nm advanced process, integrating 33.6 billion transistors, which is more than 60% higher than its predecessor Blackwell.
The Rubin platform redefines supercomputing standards through "six-chip collaboration": its super chip integrates Vera CPU and dual Rubin GPUs, and is equipped with 288GB HBM4 memory, with a bandwidth of up to 22TB/s. In terms of performance, the Rubin's FP4 inference computing power reaches 50 PFLOPS, five times that of Blackwell, while the performance per watt has increased by ten times, significantly lowering the training threshold for mixture-of-experts models (MoE). In addition, NVIDIA also disclosed the planned release of Rubin Ultra in 2027, whose NVL576 configuration is expected to push inference computing power to a new height of 15 ExaFLOPS.
