Original URL: http://www.theregister.co.uk/2008/06/18/clearspeed_700/
ClearSpeed's silicon freshness results in 41 Teraflop rack
Cited for floating point abuse
Cheaper and faster - check. ClearSpeed has hit the two major keys to winning in the silicon game with the release of its latest floating point accelerators.
The fresh CSX700 family of products offer more than double the raw performance of the previous CSX600 gear. At 192 cores, you get twice as many engines for handling floating point operations along with a pair of integrated memory controllers. In addition, the cores now run at 250MHz instead of 210MHz. And the speedy new part eats up about 12 watts of power versus 10 watts for much less silicon.
It's been about three years since ClearSpeed performed a major tweak on its flagship product. That means ClearSpeed has been trapped on IBM's 130nm manufacturing process. With the CSX700 gear, ClearSpeed has moved to IBM's 90nm manufacturing process, affording it more space for floating point cores and some power savings.
In addition, the shift freed up enough for room for ClearSpeed to throw away the FPGA it used to deliver PCI Express connections. Now, PCI Express is built-in, which again saves on power as well as cost.
The new gear should start shipping in production during the third quarter, and it could not arrive soon enough for ClearSpeed. Earlier this year, the company shed a big chunk of staff as it dealt with declining spending from financial services customers.
These days, ClearSpeed finds itself in a battle with other accelerator types, pushing things like general purpose GPUs and FPGAs that provide similar performance boosts for high performance computing applications. It used to be able to knock a lack of double-precision support among the GPGPU set, although that jab has faded now that Nvidia has joined AMD/ATI with a 64-bit chip. And the maturing FGPA accelerators from the likes of XtremeData and DRC demonstrate comparable performance per watt characteristics to ClearSpeed's accelerators.
ClearSpeed and the rest of the FPGA crowd like to stress that their products prove much more resilient in HPC set-ups. The GPGPUs lack a number of error correcting mechanisms, meaning that incorrect results can be returned, which is a major issue for customers in, say, the financial sector. So, ClearSpeed asked us to remind you all that it as ECC (error correcting code) across its whole product.
The CSX700 will slot into a couple of different products. The e710 will plug right into PCI Express lots on server boards, consuming just 25 watts while offering 4 Gigaflops per watt of performance. Meanwhile, the e720 has been built to go inside HP's c-Class blade systems, and ClearSpeed is considering doing parts for other blade makers.
There's also the Terascale box, which is a 1U server that holds 12 of the e710 accelerators. All told, this puppy provides 1.152 Teraflops of double-precision floating point muscle in a pizza box, while eating up 400 watts. It also has 24GB of DDR2 memory and a pair of PCI Express x8 connectors for host servers. Fill a whole rack with these systems, and you get a whopping 41 Teraflops of double-precision goodness.
The ClearSpeed cards used to list for more than $5,000, but that price has dropped with the new release to $3,500. ®