Nvidia Launches Blackwell Ultra GPUs Claiming 4x Performance Over...

Nvidia unveiled its next-generation Blackwell Ultra GPU architecture at the GPU Technology Conference in San Jose, claiming a fourfold increase in AI training throughput compared to the H100 Hopper generation. The flagship GB300 NVL72 system packs 72 Blackwell Ultra GPUs in a single rack-scale unit and delivers an aggregate 720 petaflops of FP8 AI performance. Nvidia CEO Jensen Huang said every major hyperscaler — including Microsoft, Amazon, Google, and Meta — had already placed Blackwell Ultra orders representing over $80 billion in combined commitments. The chips use a new NVLink 5 interconnect that doubles inter-GPU bandwidth to 3.6 terabytes per second, enabling larger models to be trained across fewer physical servers. India Tata Consultancy Services and Infosys have secured Blackwell Ultra allocations for their enterprise AI divisions.