NuAxiSemiconductor trade intelligence

H100 vs GB200 NVL72 Training Benchmarks – Power, TCO, and Reliability Analysis, Software Improvement Over Time

What happened

A benchmark analysis of over 2,000 H100 GPUs indicates current reliability and software maturity challenges are preventing large-scale training runs on NVIDIA's new GB200 NVL72 systems, with H100/H200 and Google TPUs remaining the only proven solutions for frontier AI training. The report expects significant software and efficiency gains for GB200 NVL72 by year-end, contingent on resolving reliability issues. 利多/利空/关注: 关注 - 报告未提供直接影响现货价格的供应或需求数据，但指出GB200系统可靠性问题可能延缓其大规模部署，对现有H100/H200库存构成短期中性至轻微利多支撑。