Will supercomputer speeds hit a plateau?

High-performance computers (HPCs) will not hit any speed bumps anytime soon, but escalating energy consumption by these machines is becoming a bigger issue, according to tech giants IBM and Sun Microsystems.

"The hardware speed [of supercomputers] will not reach a plateau," said Simon See, Sun Microsystems' director for Advance Computing Solution, System Practice and Global Science and Tech Network, in an e-mail interview with ZDNet Asia. "However, what might prevent effectiveness could be the software layer on the hardware."

"In order for speeds to increase in tandem with hardware capabilities, the software must not become a bottleneck," See said.

According to the Sun executive, hardware factors such as a faster CPU, more bandwidth from the CPU to memory, as well as faster and lower latency interconnectivity can contribute to the speed increase of supercomputers. Other factors include efficient file systems, as well as software with better reliability, low latency, and a good HPC library.

Jeffery Dunn, IBM Singapore's business unit executive for Deep Computing in Asia-Pacific, told ZDNet Asia in an e-mail interview that there are several methods to speed up performance of HPCs, one of which is accelerator chip technology.

"IBM believes that accelerator chips will become popular, creating an increasing variety of 'hybrid' supercomputer designs," Dunn said, citing the Cell Broadband Engine (CBE) as an example.

"Accelerator chips speed up performance by taking over some computing tasks from the main processor," he explained. "CBE is good at graphics, so a CBE accelerator would perform graphics-intensive calculations."

In September last year, IBM was selected by the U.S. Department of Energy's National Nuclear Security Administration (NNSA) to build and design the world's first supercomputer based the CBE. Dubbed Roadrunner, the machine is expected to be "capable of a sustained speed of up to 1,000 trillion calculations per second, or one petaflop".

Dunn noted that while there does not appear to be any engineering or technical hurdles that would keep supercomputing speeds from gaining, "the biggest issue is probably the ability to fund the necessary innovations".

Power up
With escalating power consumption by HPCs, the issue of energy efficiency is becoming a huge concern.

"What is key is that supercomputing can potentially take up so much power that the site where it is housed cannot allow it to function properly," See noted. "This in turn explains why our focus on innovation is also anchored to our belief in eco-friendliness."

Energy use and space requirements will become bigger issues "in this market in the near future", said Dunn.

"The energy bills for many supercomputer centers currently equal what they used to spend for the computers themselves five to seven years ago," he said. "This is why IBM designed Blue Gene to use less energy and space than other designs. We saw the energy problem when we began to design Blue Gene in 1999."

In the latest Top500 ranking of supercomputers released in Germany last month, IBM's BlueGene/L at Lawrence Livermore National Laboratory clinched the No. 1 spot, its fourth time in a row.

The biannual Top500 list measures the world's fastest supercomputers according to the Linpack benchmark, which focuses on solving linear equations.