Nvidia’s Ampere A100 GPU broke 16 world performance records in AI benchmarks

New York, 3.8.2020

The new A100 chip has more than 54 billion transistors and was developed specifically for scientific computing, data analysis and cloud graphics.

Nvidia dominated the third round of MLPerf training benchmarks for AI models from eight companies: Alibaba, Dell EMC, Fujitsu, Google, Inspur, Intel, Nvidia and the Shenzhen Institute of Advanced Technology.

The company’s A100-based supercomputer, known as Selene, led the field among commercially available systems, setting records for single node performance and scaled system performance. Selene is the fastest commercial system for AI in the U.S. with 2,048 amp A100 chips.

The Amp A100 has up to 2.5 times the performance speed of its predecessor, the Volta V100.

Google’s TPU v3 system and the smaller TPU v4 system also performed well. His TPU v3-based supercomputer had the fastest training time in several benchmarks.

The NLP model BERT (bidirectional encoder representation of transformers) and DLRM (Deep Learning Recommendation Model) were new additions to this round.