Perspective — Earlier this month, the EuroHPC JU (Joint Undertaking) reached critical mass, and it seems all EU and affiliated member states, bar the UK (unsurprisingly), have or will sign on.
The EuroHPC JU was born from a recognition that individual EU member states, and the EU as a whole, were significantly underinvesting in HPC compared to the US, China and Japan, who all have their own exascale investment and delivery strategies (NSCI, 13th 5 Year Plan, Post-K, etc).
The €1.4 billion so far earmarked comes from a repurposing of some €486 million of Horizon 2020 money, the balance coming from participating member states and commercial investment (for which the joint undertaking was necessary). In addition, a further €2.7 billion has so far been earmarked in future Digital Europe programme (2021-2027) budgets to build on the foundational work carried out by the EuroHPC project and deliver several exascale and post-exascale systems.
The intention is to deliver some pre-exascale systems and foster the creation and curation of an indigenous (or at least principally European) exascale HPC infrastructure. This ensures that key components (CPU, fabric and software stack) are based on EU derived or licensed IP and better still there is a strong inward investment component. It will also look at enlarging the pool of HPC enabled users and the expansion of existing, and creation of new, HPC competence centres to engage industry and academia alike.
Although the Commission is to be congratulated for finally listening to over twenty years of industry advice that HPC is important and central to a research-led boost to economic growth, the initiative feels like a missed opportunity.
Is €1.4 billion enough to buy in at the high-stakes game of exascale HPC?
If we look at investment in HPC, the one thing that we know beyond any measure is that it needs to be long term. Yes shiny tin with lots of flashing lights gets good PR but the real derived benefit to HPC is in the advancement of science, improved research and development. For this to work you need investment in people too. The EuroHPC project is cognisant of this and while €1.4 billion may sound like a lot of money, let’s put this into some context.
As part of the National Science Computing Initiative launched in 2015, the US administration has so far installed two exascale pathfinder systems (Summit and Sierra) with two more scheduled for late 2020 (as well as the retooled A21 in 2021). In April the DoE announced the CORAL-2 RFP which earmarks a further $1.8 billion for two, and potentially three, exascale supers to be delivered between 2021 and 2023. So let’s make that a cool $2.8 billion over five years. The NSF and the DoD, both part of the NSCI, also have their own annual HPC budgets which are in the $100’s of millions per year. Add into the mix the DARPA multi-billion, multi-year research projects, which funnel money into the indigenous HPC suppliers R&D programmes, you have a very substantial funding stream.
Japan has also long had a state-funded HPC programme. The latest iteration of which is roughly a five-year, $1 billion programme, to be spent on developing and installing the Post-K computer. Post-K is explicitly not slated to be exascale, but the goal is to deliver a 100x increase in application performance on a range of scientific codes. The CPU and fabric for the system are built on the back of decades of internal investment by the Japanese government and Fujitsu in HPC architecture and so do not represent going from a standing start, and there are commitments to buy a number of smaller systems in the $100-150 million range.
China has made funding of HPC a national priority with at least the last two 5-year plans (investing over $1 billion a year in R&D), and with the 13th 5-year plan it has explicitly made installing and operating a number of exascale installations a key priority (as well as three pre-exascale systems in the period 2018-2019). China not only has a federal funding model but it also has the political desire to ensure it is not reliant on non-indigenous technologies. China is also pushing a huge investment programme in semiconductors ($22 billion in 2014 and a further $47 billion announced in May 2018) to cut its reliance on non-indigenous devices. While this is clearly a long-term strategy, it has already paid dividends in recent years in strengthening its HPC programme.
The EU, like every other large socioeconomic block concerned about technology sovereignty, has always had a range of funding vehicles for HPC, most notably Horizon 2020. Of course the EuroHPC JU is not the only investment into pan-European HPC, with other investment streams typically amounting to around the €1 billion a year mark. This means that the figures announced by the EuroHPC, while welcome and represent a refocussing for effect; they do not represent a step change in funding.
Now let’s put HPC funding into stark relief
Intel spends approximately $12-13 billion a year on research and development in an industry segment where the total estimated investment in semiconductor R&D was in the region of $59 billion in 2017. Amazon set a new record for a US company and spent $22.6 billion on R&D in 2017. Outside IT, companies such as Volkswagen spend roughly $12-14 billion per annum on R&D (outspending other companies in the automotive industry by between 25 and 100 percent) and many pharma companies spend in the region of $10 billion per annum.
Over the two year period covered by the current EuroHPC JU (2019-2020), the EU economy as a whole will spend approximately €600 billion on R&D. German investment in R&D alone will be somewhere north of €200 billion. Will another €1.4 billion really do much to influence things? One question to ponder is if HPC is really so valuable to companies and offers such remarkable return on investment, why aren’t federal governments doing more to increase its use?
Of course some of the EuroHPC programme is being pitched as an infrastructure investment so we can put this into further perspective by looking at the scale of infrastructure investments that are being made at the same time. I’ll take the EU and the UK as examples and it’s left to the reader to find the equivalents in their own countries.
High Speed 2, the UK’s improved rail link from London to “The North,” is currently budgeted at around £56 billion (and will doubtless come in massively over budget and late as most UK infrastructure investments seem to be). The London Olympics came in at almost £9 billion. Even the Stockholm bypass project is costed at over €3 billion. Or put another way, improving European HPC infrastructure is significantly less important than putting in a bypass for a town of less than 1 million inhabitants (and by implication delivers less value)!
However, it’s hard to be convinced regarding the economic impact. With such a modest ambition, it is very unlikely that the new funding will end up anywhere except “the usual suspects”: the almost exclusively academic centres and National Champions that are already well plumbed into the European scene. Because it’s all about the people at the end of the day, generating high-tech clusters takes years and many years of sustained investment. One-off capital grants rarely succeed.
However, one area where it might make a difference is in the creation of a core set of European HPC components, to allow Europe to be less reliant on US-developed technologies.
The worldwide HPC market is not especially large (something in the region of $12 billion in 2017 for server hardware – this is separate from hardware expenditure for enterprise). To put this into context, Cray’s worldwide turnover is projected to be about $450 million in 2018. If a legal way can be found to funnel the €1 billion through a small number of European hardware companies, deploying a European CPU, interconnect, storage and software technologies, then it will be transformational for these companies and, in turn, make a very substantial difference to the HPC supplier landscape.
Chicken or the egg?
We all think that HPC is important; at least I’m making the presumption you do since you are reading this. But we somehow think that much of industry hasn’t yet quite got the message. Do we really need a series of pre-exascale machines to show them the light and make them invest in HPC?
The answer to that is probably a no. What industry needs most is people, followed closely by applications that allow them to solve their problems to a higher order of fidelity and faster. If the money that is earmarked by the EuroHPC JU for industry collaboration doesn’t tackle these issues, leading to a sustained period of investment in training and development then it will be doomed to fail when assessed against the wider context.
Yet there is an element of chicken-and-egg, in that infrastructure investments are needed to provide the free cycles given to SMEs. Without a larger cadre of HPC-enabled scientists and engineers then the expected productivity gains simply won’t materialise and nor will industry invest into more significant HPC infrastructure itself.
Industry will almost always buy what it needs from the global market. For all major powers, not just the EU, surety of HPC capability is very much a national security topic. In times of economic uncertainty and international tensions, it makes sense to maintain a degree of technical sovereignty for compute and HPC technologies, and if these cannot be wholly supported by a domestic market then these matters will take on a far more significant political overtone as they have elsewhere globally.
Europe it seems is the last major economic block to decide that it needs a longer-term strategy.
About the Author
Dairsie Latimer, Technical Advisor at Red Oak Consulting, has a somewhat eclectic background, having worked in a variety of roles on supplier side and client side across the commercial and public sectors as an consultant and software engineer. Following an early career in computer graphics, micro-architecture design and full stack software development, he has over twelve years’ specialist experience in the HPC sector, ranging from developing low-level libraries and software for novel computing architectures to porting complex HPC applications to a range of accelerators. Dairise joined Red Oak Consulting (@redoakHPC) in 2010 bringing his wealth of experience to both the business and customers.