NVIDIA’s Blackwell GB300 continues to set world records for mixture of experts (MoE) pre-training. The company reported that the GB300 remains at the forefront of AI workloads, with continuous updates enhancing its performance metrics.
Additionally, NVIDIA’s Blackwell GB200 has achieved a fourfold boost in performance per watt due to ongoing optimizations in its AI software stack. These improvements contribute significantly to the efficiency of the Blackwell GPUs.
NVIDIA’s Vera Rubin NVL72 has demonstrated a throughput of 800,000 tokens per second at 150 megawatts, surpassing the Blackwell’s performance of 80,000 tokens per second at the same wattage. This marks a notable advancement in processing capabilities.
The Vera Rubin platform is part of NVIDIA’s Extreme Co-Design initiative, which aims to provide a comprehensive stack of AI-ready hardware and software. It has been described as the most powerful AI platform available to date.
