Bristol (UK) – ClearSpeed Technology announced several new global partnerships yesterday. These partnerships will introduce their massively parallel floating-point co-processor boards to new uses bases around the globe. One thing here is clear: it’s a move toward efficiency and a greater utilization of energy resources.
ClearSpeed has announced their parallel FPU products will now be sold directly by Sun Microsystems. Sun has already installed ClearSpeed products in Warwick University’s Digital Laboratory where a complement of 24 e620 boards was showng to provide a 380% increase in 64-bit floating point throughput while only increasing the power budget 5%. ClearSpeed also announced their products will also be sold in China and Taiwan by a company called Super Sonic Technology and in South Korea by Tao Computing. The trend here is simple: more efficiency.
Running a server farm is no longer just about the general purpose CPU chugging along, cranking out data. Administrators everywhere realize the benefits of specialized compute equipment. It’s being realized that the high relative hardware cost (about $7,000 per dual-processor add-on card from ClearSpeed) is often offset by the additional performance gains in the so-called “double precision” 64-bit floating point computation relative to purchasing several more servers.
Note that the “double precision” term comes from the fact that 64-bit bits doubles the lowest standard of 32-bit floating point computational abilities (which all graphics cards are limited to in native hardware). The x86 supports three floating point formats in hardware: single, double and 80-bit “extended precision”. These provide about 8, 14 and 18 significant digits in base-10, respectively. I say about because various floating point encodings can provide additional significant digits–more on that in an upcoming feature. There are also other software-based n-bit math libraries which are slower, but do provide for an arbitrary number of significant digits.
These kinds of specialized add-on boards make a better use of electricity for what they do. They’re more energy efficient since they’re designed to do only one thing in a highly optimized manner. A general purpose CPU has to do many things (the x86 contains nearly 1,000 low-level instructions), and therefore cannot be as highly optimized for every task. In the Warwick example above, two dual-core Opteron 2218 (2.6 GHz) processors (four cores) were able to provide 20.8 Gflops of 64-bit computing per server. Each ClearSpeed card added an additional 80.64 Gflops per server, and with only 5% more power.
AMD and Intel have both invested a great deal of PR capital in conveying the notion that specialized compute abilities are not only more desirable, but they are coming directly from AMD and Intel. AMD will add specialized compute abilities to their next-gen microprocessors due out in 2009/2010. AMD reported tremendous gains in computational throughput with a fractional relative power budget for specialized operations, on the order of 40% more computing on 20% less power than general purpose CPU computing alone. Intel has announced similar findings.
The bottom line is big computing is no longer a simple science. Site admins don’t just buy a particular brand, plop it down in their server farm, write some software and begin computing. It’s about nuance, subtlety and finesse, looking not just at the capital expenditure of up-front hardware costs, but rather at compute budgets and desired processing timelines. It’s about determining what you want to do and how best to get you there for the least overall expense, and that includes long-term costs like energy and cooling.
Add-on co-processor boards like those from ClearSpeed provide a better computational value in terms of performance delivered per energy input. And with today’s advanced server farm needs, where such opportunities are realized and then seized upon, it is these products which can make the biggest difference in how efficient the installation will be.