Clearspeed Squeezes 96 GFlops Out Of 12 Watts

 

Bristol (UK) - Clearspeed has been one of the pioneers in the floating point acceleration arena, a battleground that today also includes Nvidia, AMD and soon Intel as well. Clearspeed’s new CSX700 processor-based add-in card delivers 50% more performance that its predecessor and while the device does not have the same peak performance as Nvidia’s Tesla or AMD Firestream cards, it is an extremely power efficient supercomputing board that manages to provide 41 TFlops of processing horsepower within one 16 kW server rack.

Clearspeed’s new CSX700 is one of those amazing products that are shaping the floating point accelerator segment and are not only delivering huge performance increases to supercomputers, but may impact the average desktop computer and notebook in the not too distant future as well.

The CSX700, previously code-named Callanish, includes 2 x 96 processing cores for a total of 192 cores. Manufactured in 90 nm, the 250 MHz chip includes 256 KB of SRAM, error correction hardware and integrates a DDR2 memory controller as well as a PCIe interface. Clearspeed promises that the chip will deliver 96 GFlops of peak performance compliant double precision (IEEE 754) 64-bit applications. This does not sound much, especially when compared to the latest Nvidia and AMD cards, but keep in mind that these cards only achieve about 1 TFlops in single precisions applications. Nvidia’s chip will drop to about 100 GFlops in double precision mode and consume 170 watts. The CSX700 promises to hit 96 GFlops in exchange for 12 watts. If Clearspeed can deliver on this promise, the company could have a jewel on its hands.

The company will offer the new chip as part of the e710 and e720 accelerator boards, which have the same specifications but are built into different form factors to fit different products. Both cards integrate 2 GB of RAM. There is also the CATS-700 blade, which includes 12 e710 boards for a total performance of 1.152 TFlops. Combine 18 blades within one server rack and you end up with a double precision processing horsepower of 21 TFlops and a typical power consumption of 7.2 kW. The floating point performance of such a rack would match the capability of the entire Roadrunner 2 supercomputer at the Los Alamos National Laboratory, which uses 2304 dual-core Opteron processors to deliver a peak performance of 20.3 TFlops. This system is currently listed as the 86th fastest supercomputer in the world (Top 500 list from November 2007).

Clearspeed said that the e710 and e720 boards will be priced at $3570 at launch. The street price is expected to be below $3000. There was no information on the price of the CATS-700 blade, but we were told that an estimate of $30,000 would be a good guess. The e710 and e720 are shipping now.

  • smalltime0
    ^I think it is geared to commercial sever applications than desktop gaming and server etc.
    Very interesting though, especially since the DPP comes in handy for calculating things such as protein folding. 100GFLOPS for 12W very good!
    Reply