IBM's new super computer, Roadrunner, is billed at the fastest in the world, operating at one petaflop or one thousand trillion calculations per second.

The speed demon was built for the Department of Energy's National Nuclear Security Administration to ensure the safety and reliability of the nation's nuclear weapons stockpile.

IBM said that in the past 10 years, supercomputer power has increased about 1,000 times.
Today, just three of Roadrunner's 3,456 tri-blade units have the same power as the 1998 fastest computer.Now, a complex physics calculation that will take Roadrunner one week to complete, would have taken the 1998 machine 20 years to finish.

As for the software, chalk up another one to the open source community: Roadrunner uses Red Hat Linux.

Tri-blades

The supercomputer's custom configuration uses two IBM QS22 blade servers and one IBM LS21 blade server that are combined into a specialized "tri-blade" configuration. Each tri-blade unit can run at 400 billion operations per second (400 Gigaflops). In total, Roadrunner has 3,456 tri-blades.

PDU Cabling

The system has 80 terabytes of memory and is housed in 288 refrigerator-sized, IBM BladeCenter racks taking up 6,000 square feet. Roadrunner's 10,000 connections — both Infiniband and Gigabit Ethernet — require 57 miles of fiber optic cable and weigh a whopping 500,000 lbs.

1st Stage Switch Rack Back

Roadrunner uses a first-of-a-kind design, the Cell Broadband Engine. Originally designed for video game platforms such as the Sony Playstation 3, the engine will work in conjunction with AMD's x86 processors. In total, the computer connects 6,948 dual-core AMD Opteron chips on IBM Model LS21 blade servers, in addition to 12,960 Cell engines on IBM Model QS22 blade servers.

Production Tri-blade Rear

Roadrunner's speed is roughly equivalent to the combined computing power of 100,000 of today's fastest laptop computers — users would need a stack of laptops 1.5 miles high to match Roadrunner's performance. It would also take the entire population of the earth — about six billion people — each working a handheld calculator at the rate of one second per calculation taking more than 46 years to do what Roadrunner can do in one day.

 Wiring Up CU2

Roadrunner was built, tested and benchmarked at IBM's Poughkeepsie, N.Y. plant, which is also the home of the ASCI series of

supercomputers the company built for the U.S. government in the late 1990s. IBM's site in Rochester, Minn. contributed to the

project by constructing the specialized tri-blade servers. Later this summer, IBM will load the behemoth supercomputer onto

21 tractor trailer trucks to deliver it to the Los Alamos National Lab in New Mexico.

Researchers at the Lawrence Berkeley National Lab have proposed a new model for supercomputers. The proposed model would use low-power embedded microprocessors, an approach that would overcome limitations posed by today's conventional supercomputers.

Supercomputers capable of tasks such as modeling clouds at the 1- kilometer scale are usually built by increasing the number of microprocessors at a cost of about $1 billion–plus they require about 200 megawatts of electricity to operate.

The approach proposed by Michael Wehner, Lenny Olike, and John Shalf in Towards Ultra-High Resolution models of Climate and Weather would use about 20 million embedded microprocessors at a cost of $75 million to build, require less than 4 megawatts of power, and operate at a peak performance of 200 petaflops.

To move this into reality, the Lab has signed an agreement with Tensilica to explore new design concepts for energy-efficient, high-performance scientific computer systems. The effort focuses on novel processor and systems architectures using large numbers of small processor cores, connected with optimized links, and tuned to the requirements of highly parallel applications such as climate modeling. Under the agreement, the research team will use Tensilica's Xtensa LX2 extensible processor cores as the basic building blocks in a massively parallel system design. Each processor will dissipate a few hundred milliwatts of power, yet deliver billions of floating-point operations per second and be programmable using conventional programming languages and tools.