MLPerf Training v5.1: Lambda’s NVIDIA GB300 NVL72 outperforms GB200 by 27%

Lambda's MLPerf Training v5.1 results showcase a significant performance improvement with their NVIDIA GB300 NVL72 cluster, which outperforms previous iterations like the GB200 by 27%. This advancement is attributed to the NVIDIA Blackwell Ultra architecture and enhancements in the software stack, including updates to the NVIDIA driver and CUDA, which collectively contribute to a faster training speed. The new MLPerf version replaces older models with modern workloads such as Meta's Llama 3.1 8B, aligning with the demands of contemporary AI training. Lambda's system demonstrated a 1.6× speedup over the top 64× NVIDIA B200 system and a 1.27× improvement over the best NVIDIA GB200 NVL72 submission, establishing the NVIDIA GB300 NVL72 as a leading platform for large-scale AI training. Additionally, evaluations showed a 13% average speedup using NVFP4 precision over FP8, indicating further potential for reducing training times. This positions Lambda's platform as a versatile choice for enterprise AI teams seeking a balance of performance, cost efficiency, and control.