For those who have followed supercomputing over several years, the Bull systems brand is familiar, especially in Europe. The company was acquired by Atos beginning in 2014 but from the outside, largely remains unchanged from its roots, down to maintaining the Bull name on their HPC-centric products, including its “Sequana” series of systems.
Over the years, Bull remained a well-known but small European HPC integrator with a couple of big wins and a relative handful of systems in research and academia. But the times are certainly changing, especially after important acquisitions that impact European markets where Bull has footing.
First came the sale of IBM’s x86 server business to Lenovo—a move that was expected to shake IBM off key European deals since IBM has had such a large historical presence in those HPC markets. European HPC centers are notoriously fond of their own vendors (and Lenovo is as far away as that could get), which meant few choices. IBM was native enough for a lot of centers, thus many of the European research, government, weather, and even larger enterprise HPC shops have been IBM over the years.
Cray also tended to pick up a number of wins in Europe, a trend that was gaining momentum up until the recent HPE acquisition, especially in weather and large capability-class machines. While we are not yet seeing the impact of that vendor shift in the market quite yet, we expect that in Europe it will be palpable over the new few years. This will not be because Cray or IBM are not suitable vendors or that their technologies have changed—it’s that Europe wants what it’s always wanted. Its own HPC vendor ecosystem.
There would have been no way for Bull as a standalone company to take down deals like the one that was announced today at CINECA in Italy, which we will get to in a moment. But with this system under its belt and many others that will enter the Top 500 in the next year, Atos is definitely one to watch in HPC—in Europe, at least.
In terms of system share on the Top 500 supercomputer rankings, Bull/Atos has not shown much growth, even after the acquisition. One might think a larger parent company could help Bull secure deals that would have stretched it too thin as a smaller independent, but that did not pan out over time. In 2012 through 2014, Bull maintained a 3.6% system share. In 2016, they jumped to 4%, in 2018 just 4.4% and based on the most recent rankings from the summer, they’re up to 5.2% with a total of 26 machines on the list, the largest being their #20-ranked system at French nuclear agency, CEA. As a side note, Bull as a supercomputing company would probably not exist without CEA, which needed a French, native company for its most high-value national priorities.
But overall, this relatively thin system share will change, beginning with the upcoming Top 500 list in November, and most definitely by the first bi-annual rankings in 2021.
It would seem that Atos is banking on building “baby Summit” supercomputers. By that, we mean systems that have similar GPU-dense configurations akin to the one at Oak Ridge National Lab. The difference is that instead of using IBM Power-based processors, they’ve skirted Power entirely, choosing Intel and AMD only (not that this should be a surprise). But then again, when it comes to machines like the forthcoming CINECA system, the floating point capabilities are almost exclusively driven by Nvidia’s “Ampere” A100 GPUs.
Atos is on a roll with the GPU-laden supercomputers in Europe. Last week they announced that their Bull Sequana XH2000 super would be ushered in at CSC in Finland, just before that they announced two large systems on the way, including Sweden and the UK’s “largest AI supercomputers”. They have thickened relationships with Nvidia to deliver (presumably custom) SuperPOD machines to other universities, aside from those big systems in Sweden and the UK. And in a sign that they are ready to eat Cray/HPE and IBM’s lunch in the lucrative, critical market in HPC—weather forecasting—they also took down a deal with ECMWF this year for four systems (total of 27 petaflops) and a new AI and quantum research center at the weather agency.
Today, Atos is in its strongest position yet in Europe, showing that can deliver a large, powerful system for multidisciplinary scientific and technical computing on par with the Cray/HPE and IBMs of the world. One could argue, in fact, that the “Leonardo” supercomputer announced today will usher Atos into the true “big time” in the world of HPC.
It’s hard to say where the Leonardo supercomputer will sit when it enters the Top 500 but with it is expected to reach over 200 petaflops of peak double precision performance (Nvdia is touting it as the first 10 exabyte AI supercomputer, which is based on FP16 so not necessarily the de facto when it comes to real HPC metrics).
In a call with Nvidia’s Ian Buck yesterday, the GPU maker said this will be the most powerful AI supercomputer in Europe with over 14,000 A100 GPUs, broken down across four A100 GPUs on an Atos Sequana node with a single Intel Sapphire Rapids processor with Nvidia/Mellanox HDR 200Gb/s networking.
Leonardo is procured by EuroHPC, a joint collaboration between national governments and the European Union to develop a world-class supercomputing ecosystem and exascale supercomputing in Europe, and funded by the European Commission through the Italian Ministry of University and Research.
“The EuroHPC technology roadmap for exascale in Europe is opening doors for rapid growth and innovation in HPC and AI,” said Marc Hamilton, vice president of solutions architecture and engineering at NVIDIA. “We’re working with CINECA and Atos to accelerate scientific discovery across a broad range of application domains, providing a platform to usher in the era of exascale computing.”
Nvidia appears to be the Atos partner of the year when it comes to HPC installations. What will be interesting to watch over the next year is how many HPC centers opt for the SuperPOD for more focused AI-centric computing, and how many go with the more standard HPC workload oriented architectures as provided in a Sequana. We expect it will be similar to what we’re seeing elsewhere in the AI/HPC world where nodes are GPU-heavy for both AI and accelerated scientific computing applications with a few R&D focused centers that are focused enough on AI to go full DGX or SuperPOD.
Nvidia is not in the systems business, but the partnership building several SuperPOD-based systems across Europe while maintaining their Sequana base is not just about hardware. The software and support ecosystem around these machine is massive, so while the margins might be slim in reconstructing a SuperPOD, what they make up in terms of support and solid relationships will be priceless, especially as Europe gets more ambitious than ever about its role in global supercomputing.