Nvidia Corp. said today the latest MLPerf benchmark test results prove that its latest platforms deliver the world’s fastest artificial intelligence training performance among all commercially available systems.

Nvidia’s new A100 graphics processing unit and its DGX SuperPOD system (pictured), which is a massive cluster of A100 GPUs connected with its HDR InfiniBand technology, both set eight new performance records for commercially available systems in the third annual MLPerf benchmark tests, making a total of 16 new records.

The tests were organized by MLPerf, which is an industry benchmarking group that was set up in May 2018 and is backed by companies that include Amazon.com Inc., Baidu Inc., Facebook Inc., Google LLC, Intel Corp. and Microsoft Corp., as well as Harvard and Stanford universities.

The results mark a serious improvement for Nvidia’s hardware, which previously set six records in the first MLPerf training benchmarks in December 2018, and eight records in July 2019.

Nvidia’s A100 GPU, launched in May, is the basis of the company’s third-generation DGX system, which is used to power supercomputers such as the University of Florida’s HiPerGator. The A100 chip is also available as a service on Google Cloud, targeted at companies that need the highest possible performance for data analytics, scientific computing, genomics, edge video analytics and 5G services workloads.

Nvidia said MLPerf’s latest benchmarks show that its DGX A100 system delivers a four-times performance improvement over its original DGX system that was based on its older V100 GPUs. However, it said the tests show that the older system, known as DGX-1, is now twice as fast thanks to some new software optimizations.

The latest MLPerf benchmarks included two brand new tests as well as one “substantially revised” test and Nvidia said its hardware “excelled” in all of them. For example, the A100 chip and the DGX SuperPOD system achieved the best performance in the new recommendation systems test, which is an increasingly popular workload for AI systems.

Nvidia’s hardware also achieved top scores in the Natural Language Processing category using the Bi-directional Encoder Representation from Transformers, or BERT, neural network model. It also set new records in the reinforcement learning test that uses Mini-go with a full-size 19×19 Go board. According to Nvidia, this “was the most complex test in this round involving diverse operations from game play to training.”

Nvidia said the DGX SuperPOD, which features more than 2,000 Nvidia A100 GPUs, swept every MLPerf benchmark category for at-scale performance among commercially available products.

Constellation Research Inc. analyst Holger Mueller told SiliconANGLE that MLPerf’s benchmarks are important as we’re in the middle of the race to AI, and that enterprises realize more automation is better than less.

“Covid-19 only increases the urgency, so companies are looking to platform vendors to help them with their next generation application AI loads,” Mueller said. “Today it is Nvidia’s turn, setting new records on a number of MLPerf standards. What is remarkable is that Nvidia has been able to improve its performance by four times over the last one and half years. That kind of performance increase is what is needed play in the medal ranks of AI benchmarks.”

Nvidia wasn’t alone in setting new records. Google also participated in the MLPerf benchmarks with some of its new hardware, and says they prove it has built the world’s fastest machine learning training supercomputer after setting six performance records.

Google’s latest ML training supercomputer, based on its newest Tensor Processing Unit, is four times the size of the Cloud TPU V3 Pod that set three records in the previous benchmarks last year. Made up of 4,096 TPU V3 chips and hundreds of CPU hosted machines, the system delivers a peak performance of more than 430 petaflops, Google said.

Photo: Nvidia