Over the next ten years of HPC history, the mainstream teraflop systems of today will evolve into the petaflop systems of tomorrow, while the leading-edge petaflop supercomputers will be replaced by exaflop machines. Of course, it will be up to a select few high performance computing vendors to fulfill this vision. As the most diverse player in the HPC server business, IBM has some unique advantages as it charts a path toward the exascale milestone.

One challenge for IBM will be to decide what roles its current server architectures play in the upcoming decade. Today, the company offers three basic platforms: its various flavors of x86 clusters, the Blue Gene architecture, and Power-based systems. That mix enables the company not only to own a big chunk of the overall HPC server market — 26 percent in 2008, according to IDC — but also to claim a dominant position in creating the top supercomputing systems in the world. In the latest TOP500 rankings from November 2009, IBM claimed 35 percent of the aggregate Linpack FLOPS of the machines on that list, which happens to be tops among all vendors.

According to Dave Turek, vice president for Deep Computing at IBM, even though the commercial (read mainstream) HPC market is growing faster than the cutting edge systems, it’s commitment to elite supercomputing remains strong. The rationale is that investments in high-end technology, both hardware and software, will trickle down to mid-range and low-end systems. For example, advancements in water cooling technology, which used to be a feature only in top-of-the-line machines, have spread into mainstream servers like IBM’s iDataPlex offerings. In fact, Turek expects the investments at the high end will reap greater benefits in the future than they have in the past, simply because the base of opportunity will grow more dramatically.

According to him, over the next few years, petascale computing offerings at IBM will be represented by PowerPC-based Blue Gene (/P and /Q) and Power7-based systems. “The Power side of the equation, in its various forms, will really be the centerpiece of what we do toward exascale,” says Turek. Note that PowerPC is actually an offshoot of the original Power CPUs — they have overlapping instruction sets (although the PowerPC pedigree is in low-power embedded applications, while Power CPUs have always been high-end server chips). The other interesting aspect to this is that if you discount the minor role Sparc plays, the Power and PowerPC architectures represents the last vestige of RISC CPUs in high performance computing.

At this point, IBM is much less interested in pushing x86 into multi-petaflop systems, as some of its competitors like Cray and SGI are doing, not only because of the difficulties of scaling systems based on general-purpose CPUs, but also because IBM has the luxury of driving its supercomputing aspirations with in-house technology.

Over the next few years, the Power7-based system will start to come into its own at the high end, thanks in no small part to the HPCS DARPA program which helped to drive IBM’s Power roadmap into the multi-petaflop domain. The first commercial Power7-based servers will start shipping in the first half of 2010, but its big HPC debut will be in 2011, when the “Blue Waters” supercomputer boots up at the University of Illinois at Urbana-Champaign. That machine is aiming at 10 petaflops, which is about five time the performance of ORNL’s Jaguar, the current supercomputing champ. When Blue Waters deploys it may not be the fastest supercomputer in the world, although it will surely be among the top systems.

At the recent SC09 conference, IBM was displaying some of its HPC Power7 server gear, and there was plenty to be impressed about. As expected, the Power7 implementation encompasses 8 cores and supports 4 threads per core in SMT fashion. The die contains 32 MB of embedded DRAM (EDRAM) cache, rather than static RAM (SRAM), which is faster but draws more power and requires more transistor real estate. Two DDR3 memory controllers per CPU are able to deliver 100 GB/sec of memory bandwidth (providing about 0.5 bytes per FLOP). The node includes 4 chips in a multi-chip module (MCM), 8 of which can fit in a 2U chassis, delivering about 8 teraflops of raw computing power.

As you might imagine, that much performance required a good deal of power, which is estimated to be around 800 watts per module. But since the promised performance is so high, you need far fewer servers than you would in a conventional x86-based systems to deliver comparable performance. By necessity, these HPC Power7 nodes will be water cooled, right down to the level of the chip modules themselves, greatly improving the energy efficiency.

In general, the overall design of these cutting-edge systems is focused on getting the most FLOPS/watt in the densest possible configuration. As IBM considers how to achieve three orders of magnitude improvement to reach the exaflop level in the next decade, both density and power are at the forefront of their concerns. “The energy problem, in particular, is a multi-headed hydra,” says Turek.

For years, system designers have focused on the power drawn by the CPU. Now the I/O and memory subsystems are starting to get the attention they deserve. “For exascale systems, our calculations are that the memory subsystem, left to its own devices, would be consuming on the order of 80 megawatts of power,” says Turek. According to him, the power draw by the system interconnect would be roughly the same.

The problem, of course, is that power and, to a lesser extent, space, are limited resources. They’re also resources that are not distributed evenly across the globe, which is why people are talking about deploying supercomputers in Iceland — a place where power, cooling, and real estate are rather inexpensive. However, that doesn’t help IBM or any other computer vendor very much. “From a business perspective you want to pursue a pathway that takes geography out of the question, in terms of who gets to buy and who gets to deploy,” says Turek.

It’s an open question whether Power-based systems or the Blue Gene franchise will make the trip to exascale. Like other vendors’ roadmaps, IBM’s only ventures a couple of years into the future. For the time being, the company has apparently killed the HPC Cell variant line (PowerXCell) that went into the company’s QS22 blades and powered its famous Roadrunner supercomputer. However, some Cell processor DNA will probably end up in future chips (and even in the current Power7 CPU according to a recent CNET News report) since vector-style computing seems like the shortest path to exaflops right now. And although IBM has no current plans to embrace GPGPUs in a big way, events over the next several years could always change its calculation. “Nothing stays static, for long,” concludes Turek. “That’s for sure.”