Original URL: http://www.theregister.co.uk/2006/11/16/mercury_sc06/
Humble Mercury hurls out more Cell kit
Promises big wins
SC06 Mercury Computer Systems continues to astound us with its pragmatic, humble approach to the high performance computing market.
Many of you likely only know Mercury because of its Cell processor partnership with IBM. Mercury has been selling a dual-socket Cell-based blade server that slots into IBM's Blade Center chassis, and IBM has since resold the same system as its QS20 blade.
Playing in the Cell market brings a lot of baggage, especially when you're the only real IBM Cell server partner in town. Reporters – yours included – have heaped press upon Mercury, tying the company to the glory and hype that follows the unique gaming to supercomputing Cell chip.
"We don't mind the attention, but that's not what we're all about," said Joel Radford, a Mercury VP, in an interview here at the Supercomputing conference.
Mercury has embraced the HPC market in all its current shapes and sizes. The company has FPGA gear and calls GPGPU (general purpose GPU) kit "old hat," having sealed a deal with Ziehm Imaging around the technology. Now Mercury hopes to see what ATI/AMD and Nvidia have to offer on the GPGPU front.
Mercury tells us that it's not tied to Cell at all and will focus on finding the best overall performance and performance per watt products.
That said, it has a number of new Cell systems coming in 2007.
Starting in January, customers will see a dual-socket 1U rackmount server that's aimed at HPC customers. For the first time, users will be able to create Cell-based clusters with relative ease and by using a form factor they're used to.
In addition, Mercury will ship a Cell Accelerator Board (CAB) that plugs into a PCI Express slot. University lab, media and medical imaging types are the target audience for this meaty board that chews through 210 watts – more on that later.
Then, in the middle of next year, Mercury will ship a ruggedized Cell-based server for the military.
Rivals have knocked the Cell chip for being tied to single precision floating point and for its heat consumption.
Sticking to its pragmatic roots, Mercury admits the single precision floating point limitation hurts its broader market chances for the moment. "But there are a heck of a lot of customers working with single precision floating point that can use our help now," Radford said. The Cell chip should also show a "dramatic improvement" in double precision floating point performance by mid-2008.
On the heat front, Mercury dismisses critics. The company argues that Cell boards – while toasty – show a dramatic performance per watt edge over general purpose chips, especially when you begin clustering large numbers of systems.
Moving to Cell systems can also give customers more flexibility. Researchers at Boston University, for example, had been dependent on sharing time on a large cluster. Now, however, part of the Biomedical Engineering group has moved off the cluster and onto a single Cell-based server that's all their own. Beside convenience, the group has seen a dramatic speed up as its jobs now finish in three minutes rather than three weeks.
To prove that its Cell pitch is starting to pay off, Mercury plans to announce "two to three big design wins" in the next couple of months. The wins should be for 1,000s of units, Radford said. ®