Over the past 15 years, commodity-based server computing has probably done more to mainstream HPC than any other single factor.  But specialized HPC certainly hasn't disappeared, and the market is periodically tested by those who believe that proprietary hardware is the true path to supercomputing Nirvana.

By any measure, commodity-based HPC systems -- basically, we're talking x86 Linux clusters -- dominate the industry. But because of the presence of a healthy supercomputing segment (systems over $500K), this dominance is not completely overwhelming. According to IDC's latest figures standard clusters took 64 percent of the market. At the level of the processor though, the statistics are more skewed. IDC estimates about three-quarters of HPC server revenue comes from x86-based systems, while InterSect360 Research reports fully 90 percent of systems in their most recent site survey use chips from either AMD or Intel.

In general, companies who come up with proprietary technologies (especially proprietary processors) have had limited success in this market. Often very limited. ClearSpeed and SiCortex represent two of the most recent providers of customized solutions who met untimely ends. Although ClearSpeed's accelerators offered even better performance per watt than GPGPUs, the proprietary nature of the technology kept HPC users away in droves. SiCortex and its MIPS-based clusters also offered up some very compelling performance per watt numbers, but the business couldn't reach escape velocity before rough economic times hit in 2009.

The dominance of the x86 processor has led to semi-custom designs like the Cray XT/XEs and SGI's Altix UV line. In these cases, x86 silicon is used to take advantage of the cost benefits of volume server chips (not to mention the x86-centric software ecosystem), but the design is flavored with proprietary node controllers to maximize network performance. This has proved to be an eminently successful approach from a technology standpoint, although given the lack of profits emanating from these two companies, not yet a proven business model.

There are other possible variations on the pure commodity HPC theme, one of which was evident this week in Appro's introduction of its HF1 server. In this case, the server maker incorporated overclocked x86 Xeon CPUs along with a liquid cooling system to compensate, with the idea of providing a souped-up box for high frequency trading (HFT). The servers are both expensive and warranty-challenged, but this is less of an issue for the lucrative business these servers are aimed at. It will be interesting to see if this industry-vertical customization model is a success here, and if so, if it can be replicated across other domains.

In fact, Convey Computer Corporation is aiming to do just this, in this case, with a "hybrid-core" model that employs x86 processors along with an FPGA as the co-processor. The idea is for the co-processor to be loaded with a "personality" that extends the x86 instruction set for a particular class of applications -- bioinformatics, seismic processing, data mining, financial analytics, and so on. The two-year old company has managed to grab some critical acclaim and a handful of customers, but it has yet to take the HPC world by storm.

Traveling further down the proprietary continuum, we have supercomputers like IBM's Blue Gene and its newer Power7-based HPC servers, both of which rely on custom ASICs and other hardware. Similarly, we have IBM's QS22 blade, which was incorporated into Roadrunner, the first petaflop supercomputer. That blade was based on the PowerXCell 8i Cell processor, a variant of the Cell processor used in Sony PlayStations. IBM pulled the plug on PowerXCell line when it became apparent that the market wasn't all that enthralled with Cell as an HPC accelerator.

An even more specialized supercomputer is the MDGRAPE-3 machine, developed by the RIKEN research institute in Japan. That system doesn't even pretend to be a general-purpose machine; it was designed for a single class of application: molecular dynamics. The design uses a combo of proprietary MDGRAPE-3 processors and Intel Xeon chips. There was talk of an MDGRAPE-4 a couple of years ago, but I've heard nothing about it recently

Along the same lines, is the Anton supercomputer from D.E. Shaw Research. Like MDGRAPE-3, the application target is molecular dynamics, but in this case the processing is done entirely on a customized ASIC. An article in Nature this week reported Anton recently demonstrated a protein folding simulation 100 times longer than any previous simulation -- a millisecond versus 10 microseconds. It certainly sounds like a game-changer for the protein folding folks; we just have to figure out how to put one in every lab.

Finally, there's the Green Flash project currently under development at Berkeley Lab. The idea here is to design a special-purpose supercomputer to perform climate simulations based on a much higher resolution cloud model. To be of practical use, the system would need to be about 1,000 times more powerful than supercomputers currently available, but be much more efficient in terms of power, performance, and cost. The proposed design would employ about 20 million semi-custom Tensilica Xtensa processors, cost in the neighborhood of $75 million, and draw 4 MW of power. In May they demonstrated a logical prototype of the machine by emulating the processors building blocks on an FPGA platform.

Of course, if NVIDIA has its way, system vendors will be able to create a general-purpose supercomputer with the performance characteristics approaching that of an Anton or Green Flash a few years down the road. The GPU maker's future generation processors, Kepler in 2011 and Maxwell in 2013, will be 3 and 10 times more powerful, respectively, than the current Fermi processors. Even though these future GPUs are unlikely to be as efficient as special-purpose hardware, the history of HPC suggests that designs based on commodity parts will eventually carry the day. None of which will keep people from dreaming up ever more powerful custom supercomputers.