If these numbers seem impressive, remember that this is only the tip of the iceberg. The FPGA (Field Programmable Gate Array) chip that DRC is offering is focused on cost-efficiency and versatility, but future co-processing solutions may provide even greater performance for specialized roles. Other companies, such as ClearSpeed, are working on co-processors to fit the Opteron socket, and their performance is yet unknown.
The DRC module, so named because it is a “Dynamic Reconfigurable Coprocessor,” contains an integrated PowerPC chip and runs off a Linux BIOS. It is available for around $4500 now, but Xilinx (the developer of the FPGA core) expects speeds to increase and the price to drop nearer to $3000 by next year. When you consider that it only uses 10w of power and fits into an existing Opteron server, the price seems reasonable if you're in one of the targeted industries which employ large numbers of computation-intense algorithms, such as petroleum, finance, and the life sciences. The availability of co-processors may help continue the Opteron's meteoric rise in market share, despite more competitive offerings from Intel Corp. by securing sales to these big buyers.
What, however, does this mean for the consumer? Well, right now, the chances of your next system coming with a floating point or math co-processor are pretty slim. Most people don't run applications nearly taxing enough on these functions to require anything more than the CPU's built in ability, but don't write co-processing off yet. Ageia Technologies Inc. is introducing a Physics Processing Unit at E3 specifically designed to reduce the CPU load during physics-intense games, and while it is a PCI solution, it would obviously be more efficient were it attached to the CPU with a HyperTransport link, which would provide 10x lower latency and exponentially more bandwidth. In fact, in the long term, a shift to systems where GPU, PPU, and APU units are connected to the CPU by a future revision of HyperTransport is not wholly unrealistic.