MLCommons Benchmark Tests Reveal Significant Advances in AI Training, Especially for Large Language Models

MLCommons announced its MLPerf training 3.0 benchmark results, showing major performance improvements in AI training overall and specifically for large language models.
The benchmarks included tests from Nvidia, Intel's Habana Labs, and 13 other vendors for AI training on systems with up to 384 chips.
Nvidia and Intel were the only companies to submit results for the new large language model training test, with Nvidia achieving the fastest time.
The results showed AI training speeds far outpacing Moore's Law, driven by new silicon, algorithms, software, and bigger, faster systems.
The large language model benchmark was challenging to develop but yielded insights into the systems and performance required for models like GPT-3.