ClearSpeed Ships New Math Accelerator, Inks Deal with IBM

Published: July 11, 2006

by Timothy Prickett Morgan

If all you want to do is crunch numbers, then dedicating a whole server to the task is without a doubt not necessarily the best way to accomplish that task. As has been the case many times in the past in the computer business, using a co-processor--or multiple such units--can be a much more efficient way to get the most flops in the smallest number of boxes. This is an idea that ClearSpeed Technology is trying to build a business on, and with new small form factor co-processor boards and an OEM agreement with IBM, the company is starting to get some traction with its Advance Accelerator boards.

The Advance Accelerators that were announced last year were based on a full-length PCI-X card, which limited their applicability to servers that have the room for such cards. Moreover, these cards only had one 64-bit floating point co-processor chip, dubbed the CSX600 by ClearSpeed and rated at 25 gigaflops. The CSX600 chip is actually an embedded parallel processor that contained 96 floating point processing cores. Amazingly, the chip only burns about 10 watts of juice, compared to hundreds of watts for a full server. As good as this accelerator card is, customers apparently want more. So the new half-length PCI-X card has two of these chips on it and takes up about half the space. This Advance Accelerator is rated at 50 gigaflops of raw performance and consumes about 25 watts of juice.

According to Tom Beese, ClearSpeed's chief executive officer, the new board will be appealing to a wider range of customers because of its improved performance and smaller size, and will even fit into some blade servers as well. While blade servers have not, as yet, taken off as many expected in the high performance computing and technical server market, there is a growing expectation that the total cost of ownership and easier management issues that are the key benefits of blade servers are going to start resonating with more and more HPC customers. Moreover, with faster "Woodcrest" Xeon DP processors now here and faster Opteron Rev F processors on the way in a month or so, HPC customers will be able to pack the performance they could only put in 2U servers to date, because of thermal issues, inside 1U servers. Thus, having a smaller math co-processor card is important.

ClearSpeed hails from Bristol, England, and is tightly linked to the University of Bristol. Not coincidentally, in a multi-million pound supercomputer deal that the University of Bristol announced a few weeks ago with IBM, the two are working together to build a 13 teraflops supercomputer for the school's computer science department. The super is based on a cluster of 636 of IBM's future System x Opteron-based servers, which will presumably use the Rev F Opterons that are due in early August; they were called the dual-core Opteron 2544 processors in the press release. That deal was actually done by ClusterVision, an HPC systems integrator and software supplier, which will install the hardware and has sold its ClusterVisionOS cluster management suite as part of the deal for the University of Bristol. The server nodes will access storage housed on storage arrays with 100 TB of capacity and that run IBM's General Parallel File System software. The whole shebang will run on Linux, of course.

Based on that deal, IBM and ClearSpeed have come to an agreement, and now they will be working together. Beginning in the second half of 2006 (which begins next week), IBM will be offering the Advance Accelerators in its Cluster 1350 preconfigured X64 clusters. Why IBM didn't just go all the way and support the math co-processors in its Power-based System p AIX and OpenPower Linux servers as well as its entire line of System x X64 servers is a good question. Let HPC customers buy the gear they want and the co-processors if they want them. But, when you are a small company, like ClearSpeed is, you are happy that a big company, like IBM, is helping peddle your product. And Beese is clearly happy to have IBM's endorsement, followed so closely on the heels of Sun Microsystems, which is using the ClearSpeed boards in its massive Opteron-based supercomputer cluster for the Tokyo Institute of Technology.

"This IBM deal is really important for us," explains Beese. "It is fantastic to have a company with the scale and scope of IBM take our product to market." The reason why Beese is thrilled is because ClearSpeed does not sell the co-processors directly, and its marketing plan is to work with resellers like IBM. With a price tag of around $10,000 a board, this is not a commodity product, and until volumes ramp and prices come down, it cannot be.

In addition to the IBM OEM deal, ClearSpeed also announced that Mathematica, a popular mathematics modeling program from Wolfram Research, has been certified to work with the Advance Accelerators when they are used in conjunction with Xeon or Opteron machines running Linux. While $10,000 might sound like a lot of dough, in the workstation market, a good graphics card costs that much. And the extra number crunching power--which Beese says can be as high as a factor of five on some software--is well worth it to many customers. The Mathematica software showed a speed up of around a factor of 4.5 in tests.

"Today, we are at the beginning of our adoption cycle," says Beese. "Most people in the HPOC and technical computing area can afford this." The company's goal, he says, is to ramp up volumes and drive down prices to the point where adding the cards is a "no brainer." For a few grand, I bet a few million gamers are interested, particularly if ClearSpeed can double up the chips again and get 100 gigaflops on a card.

The ClearSpeed co-processors have drivers for Red Hat and SUSE Linux as well as Windows XP running on X86 and X64 systems; support for Itanium-based machines is expected in the second half of 2006.

RELATED STORIES

ClearSpeed Ships Advance Co-Processors in Giant Sun Supercomputer

Sun, NEC, and AMD Partner for 50 Teraflops Opteron Cluster