DescriptionMLPerf Training, run by the independent, industry-wide MLCommons consortium, is the standard benchmark for comparing large-scale AI training hardware and software under audited, reproducible conditions. NVIDIA's result on the newly introduced GPT-3 175B-scale training task in the June 2023 (v3.0) round used nearly the full capacity of its Eos DGX SuperPOD system, demonstrating the growing role of purpose-built AI supercomputers in large language model training; results in this specific task category have continued to improve in later MLPerf rounds.
Formal questionNVIDIA's 'Eos' supercomputer, using 3,584 H100 GPUs, trained a GPT-3 (175-billion-parameter)-scale reference model to the target validation loss in 10.9 minutes on the MLPerf Training v3.0 benchmark, when the GPT-3 task was introduced to the suite in June 2023, the fastest submitted result at the time.
Scope auditDoes your result fit this record?
Change a parameter to stress-test whether a proposed result is still inside the published specification. This is an audit aid, not a proof checker.
This record has no editable parameters. Read the formal question and assumptions before challenging it.
Scope matches the published specificationNo parameters changed
