With plans for exascale system developments around the world firming up and actual systems due to roll out on a near continual basis starting in 2020, one thing has become quite clear: the emphasis on being able to use indigenous processor technology is a key design goal for just about every major development effort out there. It appears that gone are the days when most HPC systems were essentially large clusters that tied together the most popular, most readily available, or primarily US-based commercial processors in sufficient quantities to reach the top of the Top 500 list. Instead, exascale developers are, for a number of reasons, increasingly looking inward for their processors bases.
In China, there are at least three major exascale projects currently on the boards. Although it remains to be seen which of the pre-exascale prototypes, due to be available this year, will be selected for upgrade to exascale status, it is assured that it – or they – will be based on an indigenous Chinese processor. Indeed, each of the three Chinese prototypes use a different – and wholly unique – indigenous processor design. Likewise, recent announcements out of the EU have revealed that their overall exascale schedule will slip by at least a few years in order to take the time to design and develop the first European HPC processors and related accelerators.
In Japan, the Post-K system slated for delivery in 2022, is ostensibly an Arm-based design but in reality is decidedly a unique chip development effort by the Japanese computer maker, Fujitsu. Finally, US exascale development plans, centered primarily within the group of US Government sites charged with a host of national security applications, will once again turn to US-based processors, but perhaps with a twist. For the A21 system – the first US exascale system, slated to be installed at Argonne National Laboratory in 2021 – plans call for some form of novel chip architecture from Intel.
In the main, this trend towards more diverse and HPC-centric processors for the high-end is probably a good thing. Delivering exascale performance is no easy feat, and it is not readily obtainable simply by heaping together a sufficiently high pile of COTS[I] processors. Here, new designs that can address the demanding computational, memory, and interconnect requirements of exascale workloads likely require targeted exascale processors.
In addition, the profusion of new processor designs can only advance the state-of-the-art as a range of new features will be built, tested, and evaluated, and then either widely adopted or discarded depending on their efficacy. Finally, it can only help exascale processor development writ large that so many new HPC designers and developers from around the world – across a wide range of envisioned applications and workloads – are reaching down to the processor level to think about innovation, rather than simply using whatever the commercial sector makes available.
However, one has to wonder if there may be a downside to this trend. The most vexing concerns include:
- Needless duplication of effort.
- A more closed development environment that inhibits the sharing of successful ideas and designs.
- Perhaps most important, the negative and far reaching impact on the overall HPC software stack that will need to respond to an increasingly diverse array of esoteric instruction sets and architectural eccentricities.
It is also unclear the impact such diverse processor development will have on the overall cost of designing, manufacturing, and delivering what could be a small set of custom processors with limited market potential, at least compared with general-purpose processor sector. Indeed, the cost of exascale systems and beyond could increasingly be driven, and then limited, by the costs associated with the development of these new processor bases.
Lastly, there are concerns that center not on technical issues but on geopolitical ones. Indeed, many of these exascale processor initiatives are driven not simply as a way to ensure a steady supply of indigenous components, but rather as a way to reduce dependence on foreign supply of chips. A number of legitimate concerns serve to reinforce this perspective including: US export control policies that can in the future – as they have in the past – shut off the supply of processors critical to foreign HPC makers; increased worries with privacy and data security issues that are forcing national governments to rethink their overall data security posture; and fundamental reservations about being left behind in the global exascale race due to the uneven availability – either intentionally or unintentionally – of world-class processors.
Ultimately, this exascale processor proliferation trend is a mixed bag. Sure, the sector will get some interesting, innovative, and frankly experimental processor designs that otherwise might not have ever seen the light of day. Also, more people will be looking at more options and designs for exascale processors, and HPC makers may have the ability to pick and choose the processor – or processors – that best fit their exascale offerings. However, the sector may undergo no small degree of fragmentation as stovepipes both technical and political arise unchecked. It would be a shame if this new emphasis on indigenous processors create more problems than it solves; exascale computing is already difficult enough.
Author Bio: Bob Sorensen is vice president of Research and Technology at Hyperion Research. He also spearheads Hyperion’s nascent quantum computing focus.