Two months after their debut sweeping MLPerf inference benchmarks, NVIDIA H100 Tensor Core GPUs set world data throughout enterprise AI workloads within the business group’s newest exams of AI coaching.
Collectively, the outcomes present H100 is your best option for customers who demand utmost efficiency when creating and deploying superior AI fashions.
MLPerf is the business normal for measuring AI efficiency. It’s backed by a broad group that features Amazon, Arm, Baidu, Google, Harvard College, Intel, Meta, Microsoft, Stanford College and the College of Toronto.
In a associated MLPerf benchmark additionally launched at this time, NVIDIA A100 Tensor Core GPUs raised the bar they set final 12 months in excessive efficiency computing (HPC).
H100 GPUs (aka Hopper) raised the bar in per-accelerator efficiency in MLPerf Coaching. They delivered as much as 6.7x extra efficiency than previous-generation GPUs once they had been first submitted on MLPerf coaching. By the identical comparability, at this time’s A100 GPUs pack 2.5x extra muscle, due to advances in software program.
Due partially to its Transformer Engine, Hopper excelled in coaching the favored BERT mannequin for pure language processing. It’s among the many largest and most performance-hungry of the MLPerf AI fashions.
MLPerf provides customers the boldness to make knowledgeable shopping for selections as a result of the benchmarks cowl at this time’s hottest AI workloads — laptop imaginative and prescient, pure language processing, advice programs, reinforcement studying and extra. The exams are peer reviewed, so customers can depend on their outcomes.
A100 GPUs Hit New Peak in HPC
Within the separate suite of MLPerf HPC benchmarks, A100 GPUs swept all exams of coaching AI fashions in demanding scientific workloads run on supercomputers. The outcomes present the NVIDIA AI platform’s potential to scale to the world’s hardest technical challenges.
For instance, A100 GPUs educated AI fashions within the CosmoFlow check for astrophysics 9x sooner than one of the best outcomes two years in the past within the first spherical of MLPerf HPC. In that very same workload, the A100 additionally delivered as much as a whopping 66x extra throughput per chip than another providing.
The HPC benchmarks practice fashions for work in astrophysics, climate forecasting and molecular dynamics. They’re amongst many technical fields, like drug discovery, adopting AI to advance science.
Supercomputer facilities in Asia, Europe and the U.S. participated within the newest spherical of the MLPerf HPC exams. In its debut on the DeepCAM benchmarks, Dell Applied sciences confirmed robust outcomes utilizing NVIDIA A100 GPUs.
An Unparalleled Ecosystem
Within the enterprise AI coaching benchmarks, a complete of 11 corporations, together with the Microsoft Azure cloud service, made submissions utilizing NVIDIA A100, A30 and A40 GPUs. System makers together with ASUS, Dell Applied sciences, Fujitsu, GIGABYTE, Hewlett Packard Enterprise, Lenovo and Supermicro used a complete of 9 NVIDIA-Certified Systems for his or her submissions.
Within the newest spherical, at the very least three corporations joined NVIDIA in submitting outcomes on all eight MLPerf coaching workloads. That versatility is essential as a result of real-world purposes typically require a collection of various AI fashions.
NVIDIA companions take part in MLPerf as a result of they understand it’s a invaluable device for purchasers evaluating AI platforms and distributors.
Underneath the Hood
The NVIDIA AI platform gives a full stack from chips to programs, software program and companies. That allows steady efficiency enhancements over time.
For instance, submissions within the newest HPC exams utilized a collection of software program optimizations and strategies described in a technical article. Collectively they slashed runtime on one benchmark by 5x, to only 22 minutes from 101 minutes.
A second article describes how NVIDIA optimized its platform for the enterprise AI benchmarks. For instance, we used NVIDIA DALI to effectively load and pre-process knowledge for a pc imaginative and prescient benchmark.
All of the software program used within the exams is offered from the MLPerf repository, so anybody can get these world-class outcomes. NVIDIA constantly folds these optimizations into containers out there on NGC, a software program hub for GPU purposes.