NVIDIA has unveiled impressive benchmarks for its Blackwell Ultra architecture, showcasing transformative performance improvements that could reshape the AI infrastructure landscape. The GB300 NVL72 system demonstrates a staggering 50x increase in throughput per megawatt compared to previous Hopper GPUs, marking a significant leap forward in energy efficiency for AI workloads.

The breakthrough performance stems from NVIDIA’s expanded NVLink technology, which now connects 72 GPUs into a unified fabric delivering 130 TB/s of connectivity. This represents a massive scale-up from Hopper’s 8-chip design, combined with advanced NVFP4 precision formatting and superior rack architecture. These improvements translate to a 35x reduction in cost per million tokens, making frontier AI applications more economically viable for hyperscalers.

Particularly noteworthy is Blackwell Ultra’s optimization for agentic AI workloads, which require extensive context processing and state maintenance. The system delivers 1.5x lower cost per token and 2x faster attention processing compared to previous generations, addressing critical bottlenecks in modern AI applications. These enhancements position NVIDIA’s latest architecture as the go-to solution for next-generation AI inference tasks.

👉 Read the full article on wccftech-com.cdn.ampproject.org