Nvidia has launched efficiency knowledge for its forthcoming Hopper technology of GPUs, and the preliminary benchmarks are super.
The metrics are primarily based on MLPerf Inference v2.1, an industry-standard benchmark that analyzes the efficiency of inferencing duties utilizing a machine-learning mannequin towards new knowledge.
Nvidia claims its Hopper-based H100 Tensor Core GPUs delivered as much as 4.5x better efficiency than its earlier A100 Ampere GPUs. (Learn extra about Hopper: Nvidia unveils a brand new GPU structure designed for AI knowledge facilities) It’s a exceptional leap in only one technology. For comparability, CPU benchmarks typically develop 5% to 10% from one technology to the subsequent.
Nvidia’s efficiency leap comes with a caveat, nevertheless. The 450% enhance got here on a single benchmark; there have been a complete of six benchmarks run. The opposite benchmarks yielded at or beneath two-fold enhancements. Nonetheless, a doubling of efficiency in a single technology is spectacular.
The highest good points got here on the BERT-Massive benchmark, which measures natural-language processing of the BERT AI mannequin developed by Google and utilized in Google’s search engine, amongst different issues. Nvidia says the BERT efficiency leap is because of Hopper’s Transformer Engine, which is particularly designed to speed up coaching transformer fashions.
Ampere isn’t the one older Nvidia know-how getting trounced. The corporate additionally benchmarked Jetson AGX Orin, its Ampere-based SoC for robotics and edge methods and a substitute for the Jetson AGX Xavier processor. In these assessments, Orin ran as much as 5x sooner than Xavier whereas delivering a median of 2x higher vitality effectivity.
However I am not writing the Ampere A100 obituary simply but. Due to enhancements in Nvidia’s AI software program, it’s saying MLPerf figures for the Ampere have superior by 6x for the reason that A100 was first benchmarked two years in the past.
Orin is offered now. Hopper, which was first launched in March, is due later this yr.
Copyright © 2022 IDG Communications, Inc.
