Individuals agree: accelerated computing is energy-efficient computing.
The Nationwide Power Analysis Scientific Computing Middle (NERSC), the U.S. Division of Power’s lead facility for open science, measured outcomes throughout 4 of its key excessive efficiency computing and AI purposes.
They clocked how briskly the purposes ran and the way a lot power they consumed on CPU-only and GPU-accelerated nodes on Perlmutter, one of many world’s largest supercomputers utilizing NVIDIA GPUs.
The outcomes have been clear. Accelerated with NVIDIA A100 Tensor Core GPUs, power effectivity rose 5x on common. An software for climate forecasting logged positive aspects of 9.8x.
GPUs Save Megawatts
On a server with 4 A100 GPUs, NERSC received as much as 12x speedups over a dual-socket x86 server.
Meaning, on the identical efficiency stage, the GPU-accelerated system would devour 588 megawatt-hours much less power per 30 days than a CPU-only system. Operating the identical workload on a four-way NVIDIA A100 cloud occasion for a month, researchers may save greater than $4 million in comparison with a CPU-only occasion.
Measuring Actual-World Functions
The outcomes are vital as a result of they’re primarily based on measurements of real-world purposes, not artificial benchmarks.
The positive aspects imply that the 8,000+ scientists utilizing Perlmutter can sort out larger challenges, opening the door to extra breakthroughs.
Among the many many use instances for the greater than 7,100 A100 GPUs on Perlmutter, scientists are probing subatomic interactions to seek out new inexperienced power sources.
Advancing Science at Each Scale
The purposes NERSC examined span molecular dynamics, materials science and climate forecasting.
For instance, MILC simulates the basic forces that maintain particles collectively in an atom. It’s used to advance quantum computing, examine darkish matter and seek for the origins of the universe.
BerkeleyGW helps simulate and predict optical properties of supplies and nanostructures, a key step towards creating extra environment friendly batteries and digital units.
EXAALT, which received an 8.5x effectivity acquire on A100 GPUs, solves a basic problem in molecular dynamics. It lets researchers simulate the equal of brief movies of atomic actions relatively than the sequences of snapshots different instruments present.
The fourth software within the checks, DeepCAM, is used to detect hurricanes and atmospheric rivers in local weather knowledge. It received a 9.8x acquire in power effectivity when accelerated with A100 GPUs.
Financial savings With Accelerated Computing
The NERSC outcomes echo earlier calculations of the potential financial savings with accelerated computing. For instance, in a separate evaluation NVIDIA carried out, GPUs delivered 42x higher power effectivity on AI inference than CPUs.
Meaning switching all of the CPU-only servers operating AI worldwide to GPU-accelerated techniques may save a whopping 10 trillion watt-hours of power a 12 months. That’s like saving the power 1.4 million properties devour in a 12 months.
Accelerating the Enterprise
You don’t should be a scientist to get positive aspects in power effectivity with accelerated computing.
Pharmaceutical firms are utilizing GPU-accelerated simulation and AI to hurry the method of drug discovery. Carmakers like BMW Group are utilizing it to mannequin whole factories.
They’re among the many rising ranks of enterprises on the forefront of what NVIDIA founder and CEO Jensen Huang calls an industrial HPC revolution, fueled by accelerated computing and AI.