MLPerf HPC Benchmarks Show the Power of HPC+AI 


NVIDIA-powered techniques gained 4 of 5 tests in MLPerf HPC one., an sector benchmark for AI effectiveness on scientific programs in high efficiency computing.

They’re the most current final results from MLPerf, a set of market benchmarks for deep discovering to start with launched in Could 2018. MLPerf HPC addresses a design of computing that speeds and augments simulations on supercomputers with AI.

Modern developments in molecular dynamics, astronomy and local climate simulation all applied HPC AI to make scientific breakthroughs. It is a craze driving the adoption of exascale AI for users in both of those science and market.

What the Benchmarks Evaluate

MLPerf HPC 1. measured schooling of AI designs in 3 typical workloads for HPC centers.

  • CosmoFlow estimates details of objects in visuals from telescopes.
  • DeepCAM tests detection of hurricanes and atmospheric rivers in local climate knowledge.
  • OpenCatalyst tracks how properly programs forecast forces amongst atoms in molecules.

Every exam has two sections. A measure of how quick a procedure trains a model is referred to as sturdy scaling. Its counterpart, weak scaling, is a measure of highest technique throughput, that is, how lots of versions a technique can educate in a given time.

When compared to the best outcomes in robust scaling from previous year’s MLPerf .7 spherical, NVIDIA sent 5x better results for CosmoFlow. In DeepCAM, we delivered virtually 7x extra overall performance.

The Perlmutter Stage one process at Lawrence Berkeley National Lab led in robust scaling in the OpenCatalyst benchmark employing 512 of its six,144 NVIDIA A100 Tensor Main GPUs.

In the weak-scaling class, we led DeepCAM making use of 16 nodes for each career and 256 simultaneous work. All our checks ran on NVIDIA Selene (pictured earlier mentioned), our in-house technique and the world’s premier industrial supercomputer.

NVIDIA wins MLPerf HPC, Nov 2021
NVIDIA sent management effects in the two the pace of education a product and for each-chip efficiency.

The hottest final results exhibit another dimension of the NVIDIA AI system and its functionality leadership. It marks the eighth straight time NVIDIA sent top scores in MLPerf benchmarks that span AI coaching and inference in the details heart, the cloud and the network’s edge.

A Wide Ecosystem

Seven of the eight contributors in this round submitted results applying NVIDIA GPUs.

They contain the Jülich Supercomputing Centre in Germany, the Swiss Countrywide Supercomputing Centre and, in the U.S., the Argonne and Lawrence Berkeley Nationwide Laboratories, the National Middle for Supercomputing Applications and the Texas Superior Computing Heart.

“With the benchmark check, we have proven that our machine can unfold its likely in observe and add to maintaining Europe on the ball when it arrives to AI,” mentioned Thomas Lippert, director of the Jülich Supercomputing Centre in a website.

The MLPerf benchmarks are backed by MLCommons, an business team led by Alibaba, Google, Intel, Meta, NVIDIA and other individuals.

How We Did It

The sturdy demonstrating is the result of a mature NVIDIA AI platform that includes a entire stack of application.

In this spherical, we tuned our code with equipment available to everyone, these kinds of as NVIDIA DALI to speed up information processing and CUDA Graphs to lower small-batch latency for effectively scaling up to 1,024 or more GPUs.

We also applied NVIDIA SHARP, a vital part inside of NVIDIA MagnumIO. It supplies in-network computing to accelerate communications and offload data functions to the NVIDIA Quantum InfiniBand switch.

For a deeper dive into how we utilised these tools see our developer weblog.

All the program we applied for our submissions is out there from the MLPerf repository. We on a regular basis incorporate such code to the NGC catalog, our software package hub for pretrained AI products, field software frameworks, GPU applications and other application means.

Leave a comment

Your email address will not be published.