The semiconductor market is worth $300-billion-plus revenue per annum and Intel accounts for almost $60 billion of this total. In fact the top three companies (Intel, Samsung and TSMC) account for almost $130 billion or more than 40 percent of the total market and the next seven companies account for another $90 billion. Why is this relevant to a discussion about technological diversity in the HPC space you ask yourself?
It’s a rich man’s world
It’s a truism that no supplier has ever gotten rich from HPC — apart that is from the component suppliers such as Intel and their shareholders. For any tier 1 vendor, HPC is considered, if not quite a vanity project, certainly one that generates a halo effect rather than significant profits.
The arrival of commodity clusters in the 1990s dramatically changed the HPC market dynamics but it also inadvertently placed almost all of the real profits and control of the technical direction into the hands of a small number of commodity component manufacturers.
Now I’m not going to argue that this has been a bad thing. Intel and the cohort of top semiconductor suppliers (those who invested most heavily in process technology and foundry capacity) have ensured that Moore’s law carried on trucking. As has been oft pointed out, it became less of an observation and more of a self-sustaining marketing prophecy. The periodic restatements of the ‘law’ necessary to ensure that it remained a viable marketing tool were viewed as justified by just about everyone who was riding the wave.
One of the consequences of a move away from custom core components (such as vector processors) for HPC was that opportunities for innovation and differentiation between vendors were reduced. However a positive effect of the commoditization of HPC was that HPC was brought to the masses in a way that probably wouldn’t have happened otherwise.
If you discount IBM as an outlier in technological terms (in that they actually had four competing HPC platforms at one point as well as a foundry capability), if everyone is using the same CPUs (or at the very least the same ISA), memory, storage and to a certain extent network (InfiniBand and Ethernet) it only takes a change in one of those to potentially disrupt the market.
When the chips are down
Consider the cautionary tale of Cray, one of, if not the preeminent exponent of HPC systems engineering and integration in the market today. Excellent though the XC series and the Aries fabric clearly are, they are not a passport to huge financial reward in the HPC market. Simply put, success in supplying tier 1 capability machines of the sort that Cray excels at building, does not translate into sales into the extremely price sensitive tier 2/3 arena. Even in the good years, Cray can still struggle to deliver profit margins that would make investors heads turn.
What does this say about the HPC market as a whole? It shows that excellent engineering isn’t in itself enough. Cray themselves attempted to move down the food chain with their acquisition of Appro (2012) and in doing so started to compete more directly with the likes of HPE and Dell for cluster based HPC sales. Their more recent acquisition of the ClusterStor line from Seagate and the launch of HPC as a service under the Azure umbrella are all attempts to diversify and increase their total addressable market. The problem is that when your competitor’s revenues are literally an order of magnitude greater, simple economies of scale start to become even more relevant.
Cash flow now becomes even more critical, with purchasing power a function of how far out you can place orders and in what volumes (and probably what hedge positions you can take). Ironically, it also means that some of the systems that you are technologically well placed to build are actually too big a stretch financially without finding a deep pocketed sugar daddy (think Intel and Cray’s exascale partnership).
As an industry we now have the slightly perverse situation that, as we are entering a new era of technological innovation and diversity, as well as building things bigger and hopefully rather better, there will inevitably be a renewed phase of market consolidation.
Money makes the world go round
Now there are lots of reasons for a merger and acquisition and I’m actually willing to believe that some are definitely a meeting of minds as well as accountants. What’s also true is that in the semiconductor and computer business first mover advantage often applies. So when one of the big players at the semiconductor poker table bets big, it inevitably triggers a flurry of further M&A activity as the other players decide to follow their money or fold.
We’ve seen this recently with the ARMing of the datacentre. Softbank’s purchase of ARM in 2016 and the uptick in sentiment that ARM had finally found a rich foster parent who would invest in pushing into territory hitherto dominated by Intel encouraged a number of other moves by Intel’s semiconductor rivals.
The interplay, first between Broadcom (Avago) and Cavium (taking on the orphaned Vulcan product), then Broadcom making an opportunistic and hostile bid for Qualcomm (of course this was about way more than just the Centriq processor line) and most recently with Marvell’s bid for Cavium (again not just for ThunderX II) was interesting and instructive to watch.
From an HPC perspective it was hard to tell how much it would affect the likelihood that one or more of the ARM vendors would mount a credible challenge to Intel in the datacentre. Certainly my feeling was that Broadcom were likely to be the least sympathetic winner of the hand, but when you consider that they are playing for a share of a $60 billion pot you start to see the sort of stakes being wagered.
Which brings us to how we as an industry maintain a healthy technological diversity in the HPC market, when the reality is that only a handful of semiconductor companies have any realistic hope of challenging the current Intel hegemony.
If we look at the common denominator between the top semiconductor companies, we see that Intel and Samsung are both vertically integrated. In other words they own their own fabs and they make profits by taking their core IP (in Intel’s case the x86 architecture) into the primary consumer market and then an even more lucrative variant into the datacenter and HPC segments. Even with the eye watering capital expenditure necessary to build and equip modern fabs they are at a relative competitive advantage to those companies who have to source their devices from foundries such as TSMC, GlobalFoundries and UMC.
Qualcomm and Broadcom round out the top five semiconductor companies (both fab at TSMC) and it’s no surprise that both have been linked with ARM-derived datacenter class CPUs. Perhaps the only surprise is that Samsung appears to have sat out the hand and concentrated purely on the development of consumer space SoCs.
Of course they are far from the only companies who are looking at challenging Intel’s dominance in the datacenter but in terms of relative size, financial stability and ability to stay in the game they hold cards that none of the other players at the table do.
Even the mighty Intel is looking to grow market share via diversification (not something they have managed in the mobile space). With dominance in the datacentre (and consumer) space that they could never have expected even ten years ago, they are well aware that the next big technology wave can come along and swamp you if you’re not paying attention. Intel have made recent bets on storage (3D XPoint and NAND), the internet of things, edge and autonomous computing (Altera, Movidius and Mobileye to name but three acquisitions) and also machine and deep learning (Nervana). All of these markets are forecast to be fast growing and ultimately at least as large as the combined consumer and datacentre CPU markets.
Of course Intel will never abandon the CPU market, but recent missteps with Phi and the challenges transitioning to lower process geometries have had knock on effects in its core markets. At least some of the problems surrounding Aurora are likely to have been caused by the change in cadence for Moore’s law (assuming it’s still in ICU and not on the way to the mortuary).
There is a real diversity starting to appear, at least in the processor space, with IBM’s Power 9 and Nvidia’s Volta being stood up for the CORAL pre-exascale systems. Add to that competitive debuts for AMD’s EPYC, Cavium’s ThunderX II, Qualcomm’s Centriq, along with a range of other processors (many ARM derived), accelerators (including the reappearance of classical vector co-processors) and some real innovation in the ML/DL space and there are real competitive threats on the horizon.
Unless Intel takes a good look at some of their recent product segmentation and pricing decisions I expect that we will see a slow erosion of their numbers in 2018. With the first hyperscaler to jump ship, that trickle may become a steady stream, but until then Intel are still firmly in pole position. They market is theirs to lose but if there is a readjustment then how long can the HPC space continue to rely on what has amounted to a commodity subsidy for HPC research and development?
About the Author
Dairsie 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 a 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. He also advises clients on strategy, technology futures, HPC procurements and managing challenging technical projects.