Nvidia’s Blackwell Conquers Largest LLM Training Benchmark

Nvidia's latest Blackwell GPUs have set new performance records in the recent MLPerf benchmark competition, particularly excelling in the demanding task of pretraining Meta's Llama 3.1 403B large language model. The company's new hardware achieved top scores across all six industry-relevant machine learning tasks evaluated. AMD's newest MI325X GPU demonstrated competitive performance, matching Nvidia's previous generation H200 in a key LLM fine-tuning benchmark, indicating a narrowing gap in that specific area.

These benchmark results highlight the rapid advancements in AI hardware, crucial for developing increasingly sophisticated large language models that power many modern applications.