The Plan For Europe’s Homegrown Exascale HPC

It has been a long time since Europe has had indigenous suppliers of the core components that go into high performance computing systems. The European Union is big enough and rich enough to change that and so its member nations are funding an effort, called the European Processor Initiative, to create a homegrown processor that will be the basis of future exascale machines and that could eventually trickle down into other commercial systems.

The EPI consortium consists of 23 partners from industry, research, and educational segments, with the heavy hitters being the Bull subsidiary of Atos Technologies, E4 Computer Engineering, Extoll, Infineon, Kalray, and ST Microelectronics among the IT suppliers; Barcelona Supercomputing Center, CEA, Cineca, GENCI, Julich Forschungszentrum, ETH Zurich, Forth Institute of Computer Science, Technico Lisboa, and Universita di Pisa from the national supercomputing labs and academic labs; and BMW Group and Rolls Royce from the commercial HPC users, among others. What you don’t see on this list, not surprisingly, is any organization outside of Europe. You also don’t find the Arm Holdings, which owns the basic IP behind the processor that the Europeans will be developing and which is a subsidiary of the Softbank conglomerate.

At a recent HPC Advisory Council conference in Perth, Australia, Chris Clarkson, business development manager for Atos, gave the audience a sneak peek into the plans of the EPI. The EPI goal, concisely laid out in the following chart, is to ensure the Europe Union isn’t dependent on anyone else for its exascale processor needs. Furthermore, the member nations want to have a fully exascaled system by 2023, which is an ambitious goal, particularly since they’re looking to do it with the a semi-new (at least for HPC) Arm processor.

To us, the most important phrase in the presentation is that this homegrown processor must be “based on a solid, long-term economic model,” meaning that the EPIers aren’t just being different to be different. Rather, they are looking to build a real industry (read: a profitable one) around this new chip. More on this later.

Recent benchmarks of the Cray Arm-based Isambard supercomputer – see our initial coverage here with a follow-up set there – show that the system is mostly competitive with current Intel Skylake-based processors, but there are tradeoffs. According to Cray’s tests, the Skylake 28-core processor yields 3X more teraflops than the Cavium ThunderX2 processor used in the system. The Skylake also tops the ThunderX2 when it comes to L1 cache bandwidth (by >3X), L2 bandwidth (by 1.66X) and L3 bandwidth by a narrower 1.46 margin. The Intel processor also has 512-bit floating point vectors while the Cavium ThunderX2 only has 128-bit vectors. This might make a big difference for some applications. When it comes to memory bandwidth, the ThunderX2 shines with more than 253.4 GB/sec across two sockets compared to Skylake’s 252 GB/sec. This isn’t a huge margin, but for memory bound applications this advantage, coupled with a very significant cost difference (2X to 3X in favor of Arm), might be enough to make the difference in a buying decision.

One of the biggest advantages to Arm that is touted again and again is potential of co-design. This is a fashionable term for customizing a component to better match specific needs, like adding new HPC application friendly instructions to speed up processing. This is something that the EPI is definitely looking to exploit as we see in the slide below, and will see in other slides.

On this next slide, we get more details on the timeline and economic goals for the EPI initiative. While the first goal is to design a processor on which to build Europe’s exascale systems, there are plans to go farther than that. After the exascale systems come online, the emphasis will shift to motor vehicles, assumedly starting with Europe’s own large auto industry. BMW and Rolls Royce are part of the EPI consortium, although it’s hard to envision a self-driving Rolls Royce – not having a chauffeur sort of seems to defeat the purpose of having a Rolls.

The auto market will be considerable. More than 95 million motor vehicles were produced worldwide in 2016 (the latest year for which we could find solid figures). The number of vehicles manufactured is growing at about a 6 percent rate per year. The 2016 EU share of this production is estimated at 28 million, which is a sizeable total addressable market right off the bat. And there isn’t any reason why the EU auto Arm chip couldn’t capture business with non-EU manufacturers.

One point of emphasis for the EPI program is utilizing and developing low power processors. This is important for each of the segments they’re exploring from HPC to automobiles and everything in between. The design point for Arm fully supports this goal, given its dominance in mobile devices where battery life is crucial to success.

The Europeans have thought out their economic model and there is plenty of business they can address, including HPC, cars, space, and AI and robotics.

In his presentation, Clarkson made a point to discuss what exascale means to the European Union. It is not simply putting together a system that can hit an exaflops on Linpack, but it is all about building a system where the exascale capacity is usable with real world application performance. This is an ambitious goal, since there isn’t any code written today that can scale anywhere near exascale. (Some will argue that point, and we hope they are eventually right.)

The EPI roadmap indicates that their current plan is to couple Arm CPUs with RISC-V (pronounced risk five) accelerators to produce a first generation pre-exascale system by 2021. Tape out of the first generation processors should take place in 2019.

The second generation exascale system will come out in the middle to late 2023. The timing is approximate and a bit hazy on the chart. The first automotive proof of concept will debut in 2022 and be commercially available in the middle 2024.

Clarkson also showed an economic and organizational timeline for the first two years of EPI, starting in 2017. The consortium quickly expanded from seven original partners to the 23 who are currently signed up. The lead hardware vendor is Atos and the leading lab partner is the Barcelona Supercomputing Center.

The latest schedule is shown below, including formal milestones. EU funding, which will total €120 million ($137 million at current exchange rates), have started to flow to the project and design decisions are being made.  Atos has been elected the head of the EPI board and Philippe Notton has been installed as general manager.

In addition to the direct investment into the EPI, the EU commission has plans to provide funding of more than €1 billion ($1.15 billion) for pre-exascale systems coming in 2021 and exascale systems in 2023 or thereabouts.

 

This spending will come in the form of at least two exascale prototypes supercomputers that will be acquired in 2020 – with at least one of these systems based on EU technology and ranking in the top three HPC systems worldwide. Additional spending will arise from spending on the actual exascale systems in the 2023 time period.

An interesting wrinkle in the EPI plan is a new organization referred to as “Newco.” This is a new company that will raise at least €100 million from private sources to commercialize the intellectual property arising from EPI development efforts.

As noted on the slide, Newco will be a fabless source of advanced Arm chips, accelerators, and so forth. Newco will exclusively license EPI IP and may even contract manufacture processors or wafers for use in any segment. Royalties from IP license fees will flow back to the EPI consortium.

To us, the Newco part of the deal is critical. The EPI needs to build a compelling business case in order to attract investment from hard nosed investment and business types. If they’re successful in raising the money, then this will be a huge boost to their initiative. Getting a €100 million plus check from investors who are focused exclusively on return-on-investment would be a big stamp of approval for the EPI.

The success of the EPI is also heavily dependent on the success of their non-HPC processors and ecosystem. Building an autonomous car isn’t trivial, hardware is important, but software is the critical component – it has to work and work well in order to compete in the market and achieve regulatory approvals.

Chip volume is the key to financially supporting the exascale development of the Arm processor. The Europeans are taking an unconventional route to this by developing the exascale processor before they pursue revenue from traditional server, cloud, and auto applications. It was clear from Clarkson’s presentation that the Europeans are keen to be seen as a real player in the race to exascale and want to be considered as a solid competitor to the exascale crowns in the United States, Chinese, and Japanese for the exascale crown. Europe does not necessarily want to be first, but they want to be a close second, in our estimation.

Time will tell if Europe will be able to pull off the massive hardware and software development necessary to field a true exascale system. An interesting parallel to the EPI effort is Airbus, another multi-country European collaboration. Airbus has managed to succeed over time versus tough competition from Boeing, selling more single aisle planes than the American company and, more importantly, doing profitable business in a cutthroat industry. EPI is an Airbus-like collaboration addressing even bigger markets, can they engineer the same success?

You can see the entire presentation by Clarkson here.

Sign up to our Newsletter

Featuring highlights, analysis, and stories from the week directly from us to your inbox with nothing in between.
Subscribe now

Be the first to comment

Leave a Reply

Your email address will not be published.


*


This site uses Akismet to reduce spam. Learn how your comment data is processed.