Why supercomputing holds the answer to a greener tomorrow
Sophie Proust
Chief Technology Officer, Atos
Jean-Pierre Panziera
Chief Technology Officer for High Performance Computing, Atos
Posted on: 18 November 2019
With the finishing line in the ‘Race to Exascale’ in sight, we see the development of High-Performance Computing (HPC) solutions for real-world use cases, radically cutting both costs and emissions
“We’re developing supercomputers for real-world use cases,” says Sophie Proust, Group Chief Technology Officer and chief advocate of Atos’ sustainability credentials in the supercomputing race. Atos experts with more than a decade of experience in the HPC space aren’t developing supercomputers optimized for benchmarks and ad rankings, they are focusing on real-life applications
But with sustainability high on almost everyone’s agenda these days, will supercomputers ever really fit with low carbon ambitions? After all, scale means you're going to have more servers, nodes, network and storage, each requiring more energy.
Innovations enhance sustainability credentials
There’s no doubt that the ‘Race to Exascale’ requires pioneering efforts for immense technological advancements that will certainly change the world we live in, just as the ‘Space Race’ did in the 20th century.
Supercomputers also get super hot, and keeping them cool will require inordinately large amounts of electricity.“You can use as much power cooling a system as you do running it,” reveals Jean-Pierre Panziera, CTO for HPC at Atos and Chairman of ETP4HPC, the European Technology Platform for High Performance Computing, an industry-led think-tank promoting European HPC research and innovation to support Europe’s competitiveness. “Every megawatt amounts to one million Euros a year on average.”
If that thought is making you sweat, Atos has taken the heat off. Its liquid cooling uses warm water – up to 40°C – to keep systems up and running at the right temperature. Jean-Pierre Panziera reveals how Atos “spotted the problem and started working on resolving it ten years ago.” Liquid cooling, he points out, is now “a very mature technology that can be deployed in a very large configuration in a very efficient way.”
Liquid cooling is just one such example. Without it, Exascale couldn’t realistically be applied to everyday uses in healthcare, defense, maintenance, transport – and virtually every industry you can think of.
Co-creation drives innovation
Supercomputing is a booming area with a lot of new technology, and a fast-changing landscape, as Sophie Proust confirms: “HP has acquired Cray. NVIDIA has acquired Mellanox.” Seeing – and understanding – the evolving landscape is clearly important.
As is co-design, which is another critical aspect in the ‘Race to Exascale.’ While historically France’s CEA (Atomic Energy Commission) was Atos’s main partner for HPC development when it all began back in 2004 at Bull. “But now it's expanding,” reveals Jean-Pierre Panziera. “For example, we have now taken a co-design approach with the Juelich supercomputing centre in Germany to developing our next GPU accelerator modules.”
And hand-in-hand with co-creation comes co-design. It takes a deep understanding of how hardware is evolving for Atos to be able to answer critical design questions. Should I include a quantum acceleration? Would I use a GPU? Would I use a field-programmable gate array (FPGA)? Would I use artificial intelligence (AI) in some of my applications? Would I offload some of my processing in the cloud?
Grasping the merging and converging
Understanding how technologies such as artificial intelligence and the cloud are evolving is particularly critical. Why? Because they are starting to merge (and converge) with supercomputing technologies.
Firstly, bringing HPC and AI together means problems can be solved more sustainably. Solving a supercomputing problem can take months of processing numerical simulations. Bring in AI to solve part of the problem for you in the first instance before running the numerical simulation. “It’s much faster. You use less energy,” advocates Sophie Proust. This new approach will also give you the means to handle problems that were previously out of reach.
And secondly, merge HPC and the cloud to compute some problems “at the right place for the right usage.” It helps with sustainability too!
But it’s not just about merging the technologies; it’s also about convergence, as Jean-Pierre Panziera explains: “AI was something that you could only do on a small scale. At a larger scale, you need more power and you need large infrastructures that are strongly coupled, which is something that is the mainstream of HPC. And the other side of this convergence is that people have already started using AI in HPC applications. But AI can also be used to improve the administration and maintenance of the data center, what we call the Cognitive Data Center.”
A home-grown industry for Europe
We’ve talked about a wide range of companies and technologies here, but they’re just the tip of the iceberg in the race; the contest is truly global. The US has a national program, as do Japan and China. Likewise, Europe has its own ambition to develop a home-grown HPC industry.
Jean-Pierre Panziera is very familiar with the European High-Perfomance Computing Joint Undertaking (EuroHPC): “EuroHPC is one of the first steps toward developing local HPC know-how and technologies, already federating all EU Member States into a single program. Switzerland, Norway and Turkey have also joined the program. And we are part of the equation, proud to be contributing to the development of advanced technologies in this space.”
Europe’s open market approach is, undoubtedly, a challenge. “The HPC market is seen around the world as strategic,” explains Jean-Pierre Panziera. “It's something you need to own, master, control, and some countries protect it with their national programs.”
How close is the finishing line?
Before discussing the timeline, we have to acknowledge that Exascale is a moving target where “once you’ve solved one step, the next one comes along.”
Currently, the systems in the field, Jean-Pierre Panziera clarified, are what’s known at a European level as ‘pre-Exascale.’ From a European perspective, there is a series of calls under EuroHPC for the procurement and operation of three pre-Exascale systems, to be installed toward the end of 2020.
But Exascale systems that are 10x or even 100x more powerful than the Petascale systems we have today, still have to be developed. “It’s the next milestone for HPC and what everyone in the HPC world is aiming for,” states Jean-Pierre Panziera. “We should be reaching that in around 2022 or 2023 for our European customers. Some earlier systems may be installed in the US in 2021 or early 2022.”
EuroHPC, through its Research and Innovation program, is also advancing the co-design of European HPC technology and the development of the first generation of a European low-power microprocessor technology as well as fostering applications, skills development and the wide use of HPC.
Atos is expanding its footprint too, from the early days in France to across Europe today, even “seeing interest from customers around the world.” Last year Atos signed a partnership agreement with India to develop their own HPC industry and know-how as part of India's National Supercomputing Mission.
Exascale will eventually arrive. Sophie Proust is proud that Atos is “not doing the technology to be the first in the ranking” but instead “working on technology that really is usable.” Technology for real applications, for real uses, and for the real world, where reducing power consumption is essential for a sustainable future.