Another supercomputing show, another performance record, courtesy this week of Oak Ridge National Laboratory’s Summit system, the first to smash through the 100-petaflop barrier.
But behind that headline at Frankfurt’s International Supercomputing show is a broader story about how the tectonic plates of the supercomputing world are shifting.
As Moore’s law continues to slow, accelerated computing clearly emerged at ISC as the booster rocket that will soon propel us into the age of exascale computing. Consider: Most of the new processing power on the just-released Top500 list comes from GPUs, which provide 95 percent of Summit’s flops.
And NVIDIA Volta Tesla Core GPUs are providing that power, enabling multi-precision computing that fuses the highly precise calculations that tackle the challenges of HPC with the highly efficient processing required for deep learning.
Indeed, the Top500 list shows that five of the world’s top seven supercomputers are now GPU powered, including the top systems in the U.S., Europe and Japan.
Summiting with Accelerated Computing
Those who track the big twice-annual supercomputing shows have seen accelerated computing on the rise over recent years. But at ISC18, it broke past the tipping point.
Summit is clearly the most potent example. Powered by 27,648 Volta Tensor Core GPU, it was measured at 122 petaflops of double precision-performance. Its performance each second is equivalent to the Earth’s entire population doing one calculation a second for an entire year.
And it’s AI performance speed is even more dazzling at 3 exaops. That’s like the entire Earth’s population doing one calculation a second for 15 years.
Mean and Lean
Multi-precision computing opens up new worlds of possibility. But that would be of limited utility if GPUs didn’t also offer extraordinary efficiency.
GPUs now power 17 of top 20 greenest systems in the world, according to the new Green500 list. Summit isn’t only the world’s fastest, it’s also the world’s most efficient system within the newly established “Level 3” category, the most stringent of the levels in Green500 list.
GPU’s have helped improve power efficiency by 50x in the past 10 years for leadership class supercomputers at Oak Ridge national labs, going from CPU-only Jaguar to GPU-accelerated Titan and Summit.
And all this is just a start. Achieving exascale will require even more breakthroughs in power efficiency. With the average efficiency of systems in the Green 500 list, powering exascale would take over 300 megawatts of energy, equivalent to the power requirements of 250,000 U.S. homes. Exascale requires 10X higher efficiency to operate in 30 megawatts.
GPUs have gotten Summit halfway toward this ambitious goal, and offer a clear path forward to efficient exascale by 2021.
Cutting Through the Knots
The once-unimagined processing capability of the latest top systems makes it possible for today’s generation of researchers to address some of science’s knottiest challenges.
Take, for example, genetics. GPU computing power can unlock such puzzles as the link between the human genome’s billions of AGCT DNA pairs and devastating diseases like Parkinson’s and Alzheimer’s. Already, Summit’s making headway in combing through an individual’s genes to determine sensitivities to opioid addiction – one of the leading causes of death in the U.S.
Or take materials. Superconductive materials can be used to develop powerful scientific magnets for MRI equipment, particle accelerators, or magnetic fusion devices. Today’s versions, however, are brittle, hard to manufacture and only work at very low temperatures. Summit is helping simulate and discover new superconducting materials with metal-like properties that can operate at room temperature.
Or take cancer research. A key to combating cancer is developing tools that can automatically extract, analyze, and sort health data to reveal previously hidden relationships between such disease factors as genes, biological markers, and environment. Paired with unstructured data, like text-based reports and medical images, deep learning algorithms scaled on Summit will help provide medical researchers with a comprehensive view of the entire U.S. cancer population at a level of detail typically obtained only for clinical trial patients.
Just Getting Going
We see this as just the beginning for accelerated computing.
Every country is racing to build exascale systems. Peek at the Top500 list of 2025 and you’ll likely see over a dozen of such systems, with multi-precision accelerated computing the platform of choice. By comparison, all the systems added together on this week’s new Top500 list barely achieved an exaflop of total computing. This speaks to the massive opportunity ahead.
One of the great appeals of accelerated computing is that it’s full-stack innovation — from the architecture, through to the system, acceleration stack, developers, as well as semiconductor process. That’s important because, with the end of Moore’s law, there are no automatic performance gains.
At NVIDIA, we’ve been investing in accelerating the full HPC stack for more than a decade.
When we started with the first CUDA-capable GPU, it could run exactly zero applications. An entire universe of applications, algorithms, libraries, tools, compilers, operating systems, and system design needed to re-designed for a new accelerated world. It’s easy to build a chip that stamps out math processors; making those processors usable and programmable by the world’s HPC developers takes extraordinary innovation on the entire stack.
As a result, more than 550 HPC and AI applications are GPU-accelerated, including the top 15 applications and all AI frameworks. The number of developers working on this is now close to a million, up 10x in the past five years. And with the latest HPC containers on our NGC container registry, HPC users can now simply click, download, and run the latest GPU accelerated apps on their systems or in the Tensor Core GPU powered cloud.
Looking Around the Bend
Now that we’re barreling down the accelerated computing straight-away, some of us are looking around the next bend to quantum computing, which uses quantum bits, or “qubits” instead of 1s and 0s to handle information.
These theories are deeply intriguing. At some point in the future, there may be killer apps that run on quantum computers, particularly in the area of cryptography or quantum chemistry, taking advantage of extraordinary processing power that draws exceptionally little power.
But for the foreseeable future, accelerated computing’s momentum appears unstoppable. We are committed to continuing to innovate in HPC, putting the promise of exascale — and all that it holds for science — within our grasp.