IDC’s HPC team — Earl Joseph, Steve Conway, Richard Walsh, Jie Wu and Dan Lee – believe that current market conditions and technology trends support the following predictions:
1. The HPC Market Will Maintain Strong Growth
In 2008, HPC server market revenue will vault past $12 billion (2006: $10 billion). Clusters will make up more than 70 percent of this total, and cluster sales will shift even more notably toward blades. Systems in the Workgroup (below $50,000) and Departmental ($50,000 to $250,000) segments will account for more than 60 percent of server revenue — about 5 of every 8 dollars spent. Worldwide year-over-year server growth will average just under 9 percent. Average worldwide growth for the HPC storage market will be higher than for servers, about 11 percent. Total revenue for HPC servers, storage and services will surpass $19 billion (not including costs for software applications, staffing, facilities and power).
2. The “Petaflop Club” Will Gain Its First Member(s)
The number of nations pursuing the prestigious (albeit arbitrary) petaflops milestone is growing apace. “Petaflop Club” contenders need not be global superpowers: the Swiss Federal Assembly may fund a system. They need not be located in the U.S. or Japan: multiple European nations, as well as China and India, are also in the hunt. In 2008, it’s likely that one or more systems will attain Linpack petaflops performance and the attendant rights to chest-thumping publicity. This will pave the way for what Horst Simon (Lawrence Berkeley National Laboratory) has dubbed The Petaflop Era and for the tougher goals of sustaining petaflops performance on an “embarrassingly parallel” 64-bit real world application, and later on a more challenging code. The U.S. and Japanese petascale programs currently appear to be best positioned for this kind of heavy lifting.
3. The High End Will Get More Competitive
While the market for HPC systems priced above $5,000,000 has been declining, the new IDC Supercomputers category (systems that sell for over $500,000) continues to show growth and offers unique opportunities for mindshare and technology innovation that can benefit the larger HPC market and vendors’ overall businesses. The top 10 spots on the November 2007 Top500 list hint at the increasing geographic and vendor diversity we can expect to see in the most expensive systems category. In the November 2007 top 10, Germany, India and Sweden were represented alongside the U.S., and IBM, Cray and SGI were joined by HP. Will Borat deliver the keynote at SC09?
4. Lower End Stereotypes Will Start to Dissolve
The Workgroup and Departmental segments are the biggest growth engines for the HPC market, but far less is known about these users than their higher-end counterparts. IDC research in 2007 revealed that the settings for low-end HPC systems are more diverse than was previously believed, varying from startups to multibillion-dollar industry leaders (some of which have more than 1,000 workgroup systems).
The research also showed that although these sites generally are not doing breakthrough science and engineering, their workloads typically include both capacity work and innovation-targeted capability jobs. The financial attractiveness of the lower-end segments, especially to big hardware systems vendors and mega-firms such as Intel, AMD and Microsoft, will increasingly cause light to be shed on these users and dispel more myths in the process. (IDC plans to ramp up its 2007 lower-end research aggressively in 2008.)
5. Explosive Low End Growth Will Drive “Ease-of-Everything” Solutions
Much remains to be learned about the low end, especially the white-hot sub-$50,000 Workgroup technical server segment. But this much is already certain: these users typically require ease-of-everything — purchasing, installation, operation (including software compatibility and performance), and upgrading. Whether in small engineering services firms migrating to technical servers for the first time or departments within huge organizations with years of HPC datacenter experience, these users typically lack easy access to HPC-experienced IT personnel.
In 2007, HPC vendors began fielding new systems with starting prices in the $15,000 to $25,000 range (e.g., SiCortex Catapult, HP Cluster Platform Express, IBM, etc.) and pre-tested reference architectures (e.g., Intel Cluster Ready and various entrants from HP and IBM) aimed at addressing ease-of-everything Workgroup requirements. And the continuing growth of Platform Computing (a partner with Intel and Dell in the Intel Cluster Ready architecture) attests to the value of proven middleware. IDC predicts that this trickle of new Workgroup offerings will turn into a forceful stream in 2008.
6. 2008 Will Be a Breakout Year For Power Efficiency
Within the triad of HPC environmental challenges (power, cooling and facilities space), one — electrical power — is fast becoming the most important. 2008 will see more emphasis on power-efficient HPC systems and more RFPs specifying power limitations. This trend in turn will support the increased use of accelerators, which for certain problems can exploit more parallelism at lower clock periods to deliver higher performance. With energy prices increasing and no lasting relief likely, 2008 should also speed the trend toward locating the most powerful computing facilities in areas with comparatively cheap energy (commercial example: Google; HPC example: Oak Ridge National Laboratory). This trend could eventually cause top scientific talent to concentrate in the same areas to take advantage of the best computational resources.
7. 2008 Will Further Define Accelerator Opportunities (and the x86 Turf)
With their backs increasingly against the x86 single-core frequency wall, HPC vendors and users will make progress in scaling application performance on multicore processors, but in many cases this progress won’t be enough. That means accelerators will play an increasingly important role in 2008 and for the near-term future. The key unknown questions are which accelerators to use and how large an economic opportunity HPC accelerators represent. If the opportunity is only moderate, connections mediated by HyperTransport and PCI Express may be enough; if it’s larger, microprocessor vendors may offer “chips with personality” that embed the accelerators directly on the die.
In 2007, GPUs, FPGAs, ClearSpeed, IBM Cell and vector processors (Cray, NEC) all made progress. IDC expects heterogeneous processing to gain traction in 2008, with GPUs currently in the limelight. In the x86 camp, 2007 closed with Intel subjecting AMD to a technology attack while AMD maintained near-silence in defending its HPC franchise. Will AMD strike back in 2008, at least in a marketing context? Will accelerators be part of AMD’s counterattack?
8. Software Licensing Costs Will Approach Crisis Proportions
IDC research showed that licensing costs started to reach crisis proportions for many users in 2007, driven by key software vendors continuing to use pricing models based on cores and choosing not to adapt their pricing models to match the rampant growth in multicore hardware parallelism. Some ISVs have introduced promising alternative models, including token-based approaches (e.g., Altair, Allinea), but more time is needed to gauge the extent of their success. Meanwhile, traditional ISV pricing models will act as a brake on scaling up hardware systems and applications performance at budget-constrained end user sites.
9. Interconnects Will Begin Moving to Fiber
In 2008, watch for HPC cluster compute and storage interconnect media to begin transitioning to fiber. Luxtera signaled its readiness to enter the production stage in Q4 2007 with its QDR-ready InfiniBand product and the hire of business ramp-up expert Greg Young. DDR 20 Gbit/second-capable optical cables from Intel and others are on the market now and offer important advantages over 24 AWG copper (greater range, higher bend radius, higher bandwidth, and fewer errors). In the 2008 interconnect battle, IDC expects InfiniBand to gain momentum as 10 GigE takes time to find its way into the market.
10. Storage and Compute Networks Will Increasingly Integrate
As larger configurations with multicore processors proliferate, storage lags further and further behind HPC compute power. This means that the need for storage capacity and capability is greater than ever. IDC expects the market for HPC storage to reach about $4.8 billion in 2008 (2006: $3.8 billion). Large commercial storage vendors such as EMC and NetApp are focusing more on HPC. Mellanox is also focusing more on the HPC space as DDR Infiniband takes off and stimulates HPC storage and compute networks to increasingly integrate.