This week at the International Supercomputing Conference (ISC) in Dresden, Germany, IBM unveiled its next-generation Blue Gene architecture — Blue Gene/P. The new model is intended for users looking for petaflop-level computing and beyond. Like its Blue Gene/L predecessor, Blue Gene/P is targeted for big science applications and the very highest end of commercial HPC. According to IBM, Blue Gene/P is two and a half times more powerful than the Blue Gene/L generation and requires only slightly more power. A relatively modest two-rack Blue Gene/P configuration that IBM deployed in-house ended up as number 31 on the new Top500 list announced this week.

The previous-generation Blue Gene/L machines represent some of the fastest systems in the world. The Lawrence Livermore system currently holds the top spot on the Top500 list and a number of other Blue Gene/L installations are scattered throughout the list. But this is the end of the line for Blue Gene/L. IBM manufacturing will now switch over to the P line. Blue Gene/L purchases currently in the pipeline represent the last machines of the first generation.

The Blue Gene/P architecture is based on a quad-core PowerPC 450 ASIC chip, where each chip is capable of 13.6 gigaflops. A compute node includes the ASIC chip along with 2 GB of SDRAM DDR2 memory. Thirty-two of these nodes are aggregated onto a board and 32 of the boards are placed in a 6-foot high rack. The result is a 4096-core rack, which provides 13.9 teraflops (peak) of processing power. This represents the smallest Blue Gene/P system you can buy.

Although the original Blue Gene/L was dual-core, it did not implement cache coherency in the hardware. By contrast, Blue Gene/P is designed around cache coherent quad-core chips, so they can be treated as SMP nodes in the same manner as any multicore-based commodity cluster. This makes the new Blue Gene more suitable for multithreaded workloads based on standard software technologies like OpenMP.

Compared to Blue Gene/L, the new generation uses slightly faster PowerPC processors (850 MHz versus 700 MHz) and twice as many cores per chip (4 versus 2). L3 cache has been doubled from 4 MB to 8 MB and main memory per compute node has been quadrupled from 512 MB to 2 GB. Main memory bandwidth has also increased — from 5.6 to 13.6 GB/sec. In addition, the 3-D Torus and Tree networks have been upgraded, essentially more than doubling the bandwidth and cutting latencies in half. The increased capabilities provide a 2.4x increase in performance over Blue Gene/L, using roughly the same floor space and slightly more power.

A single 13.9-teraflop Blue Gene/P rack draws just 40 kilowatts, yielding 0.35 gigaflops/watt — possibly the best performance/watt metric of any general-purpose computing system on the planet. SiCortex’s MIPS-based cluster systems come close at around 0.32 gigaflops/watt. For comparison, Blue Gene/L offers a lower, but very respectable, 0.23 gigaflops/watt. Most x86-based high performance computing systems are an order of magnitude lower than that. As users build Blue Gene/P systems that scale to hundreds of teraflops and beyond, power efficiencies become even more critical.

And while not every customer will use Blue Gene/P to build petaflop systems, IBM anticipates at least one customer will put enough Blue Gene/P racks together to reach a sustained (Linpack) petaflop as early as next year. Apparently IBM has a few prospects that are considering purchasing the 80 or so Blue Gene/P racks required to build a such a machine. The architecture is actually designed to scale up 256 racks, which would come close to three Linpack petaflops. However, there are few customers who would know what to do with such power, and the cost would probably be prohibitive even for that select group. IBM realizes that, although there are many HPC customers with computational problems bigger than their machines, there are only so many organizations that have the right combination of money, workload, and software experience that’s required to take advantage of machines like Blue Gene/P.

In any case, the Blue Gene/P sales pipeline is already filling up. The U.S. Dept. of Energy’s Argonne National Laboratory, Argonne, Ill., will deploy the first Blue Gene/P system this fall. Argonne’s initial Blue Gene/P system will be a 114-teraflop machine, and the lab is on track to eventually install about half a petaflop. Argonne currently has a Blue Gene/L system and will continue to operate that machine through at least 2008.

Explaining the lab’s motivation to increase their Blue Gene investment, Ray Bair, Division Director for the Argonne Leadership Class Facility said: “Blue Gene has been a resounding success for scientific computing since its inception, both for DOE’s INCITE program at Argonne National Laboratory and in diverse science programs at institutions around the world. The breadth and scale of science problems that can be addressed with Blue Gene was another important factor. IBM designed Blue Gene/P with petascale scientific computing in mind, making performance and functionality improvements from top to bottom while preserving Blue Gene’s extraordinary balance.”

Other installations are being planned as well. In Germany, the Max Planck Society and Forschungszentrum Jülich are scheduled to begin installing Blue Gene/P systems in late 2007. Other Blue Gene/P deployments are being planned by Stony Brook University and Brookhaven National Laboratory in Upton, N.Y., and the Science and Technology Facilities Council, Daresbury Laboratory in Cheshire, England.

Since the public sector is the principal source of the money for capability-class supercomputers, Blue Gene systems tend to live almost exclusively in government labs and facilities. IBM has courted Wall Street, but has not closed any Blue Gene accounts there. With the increased emphasis on power and cooling costs, IBM is hoping that a tipping point will occur at some point and commercial entities will consider Blue Gene supercomputers as cost-effective alternatives to large HPC cluster systems.

What IBM is really counting on is that a good proportion of the installed Blue Gene/L base will upgrade to Blue Gene/P. Applications should port rather easily, requiring only a recompilation. Unfortunately, the two architectures won’t interoperate; a Blue Gene/P system won’t be able to bolt onto a Blue Gene/L rack to accelerate the original system. IBM does, however, offer a trade-in program for customers looking to retire their older models and get some credit against a Blue Gene/P purchase.

Because of the scale of the architecture and the extended lifetimes of these types of supercomputers, IBM put a lot of thought into reliability and system robustness. Efficient cooling design, soldered memory, and a low number of moving parts supports a low mean time between failure (MTBF) rate. From the feedback they received from Lawrence Livermore National Laboratory, it turned out that the lab’s Blue Gene/L system had an order of magnitude better MTBF than commodity-based systems installed there. When added to the cost savings realized from the system’s power efficiency, IBM offers a fairly compelling TCO story.

This message is tougher to sell in the commercial HPC space, where customers are still very sensitive to initial acquisition costs. According to Herb Schultz, Deep Computing Marketing Manager, IBM is aiming for around ten cents per megaflop for the new Blue Gene systems, which he feels is price competitive with other non-discounted HPC systems in the industry. But with the smallest installation being a 14-teraflop system, customers are looking at $1.4 million to join the Blue Gene/P club.

“If we can get customers to look more broadly at the overall system costs — the power and cooling bill over three or four years and the costs associated with system downtime — we think Blue Gene/P looks really good,” said Schultz. “So we’re trying to get people to look beyond the initial acquisition cost and focus on the total cost of operating the machine over its lifetime.”