ClearSpeed expands math accelerator line

ClearSpeed has added a second option to its line of add-in cards that boost a computer's computational power with a special-purpose processor.

The company's first Advance X620 card plugged into servers' PCI-X slots, the last of an older generation of input/output technology. The newer model, called the e620, uses the PCI Express technology common on mainstream servers and also in PCs and workstations.

The new model costs the same--$8,000, before volume discounts--and offers about the same performance, said Peter Ffoulkes, a marketing director at ClearSpeed. That price may seem steep, but the company argues that it's cheaper than the alternative--buying extra servers--and the power consumption is lower, Ffoulkes said.

However, the new model arrived later than the start-up had hoped. "We are a little bit behind on that one," Ffoulkes said.

ClearSpeed is trying to tap into two trends in high-performance computing. The first is the use of compute clusters, groups of interconnected servers that collectively tackle calculation challenges such as processing three-dimensional oil reservoir data or simulating the interactions of drugs and human cells.

The second trend is the use of special-purpose hardware to accelerate such clusters. That type of hardware--which also includes the Cell processor from IBM, Toshiba and Sony, repurposed graphics chips from Nvidia and AMD's ATI division, and field-programmable gate array (FPGA) chips that can be configured to run all sorts of tasks--can boost performance, but software must be adapted to take advantage of them.

To use ClearSpeed, customers' software must tap into prewritten libraries from the company rather than the typical libraries that the software otherwise would use. ClearSpeed also announced on Tuesday that its libraries support Microsoft Windows, not just Linux. Microsoft is trying to coax Windows into the compute cluster market but so far has gained only a toehold.

ClearSpeed also added support for some new types of calculations, including "vector" operations and calculations requiring random number generation.

The company's tests show that four Hewlett-Packard DL380 G5 servers with dual-core Xeon 5300 processors can perform 115 billion calculations per second, or gigaflops, on the widely used Linpack speed test. Adding an X620 card to each of the servers boosts performance to 251 gigaflops.

Reaching that level without the cards requires nine servers, ClearSpeed said. And average power increases much less: from 1,722 watts with the four-server cluster to 1,750 with the four ClearSpeed cards added, compared with 3,875 for the nine servers.