One of the most competitive markets in computing is servers, with companies always looking for more efficient processing to accommodate more complex simulations or decrease operating costs. For years, the X86 segment of the server market was controlled by Intel's Xeon offerings, but in recent years, a new contender has emerged, AMD's Opteron.
The Opteron has taken the market by storm; in the first quarter of 2006, it accounted for 22.1% of X86 server sales, a 26% increase on the previous quarter, and a 254% increase on the previous year. A large part of the Opteron's success has to be put down to HyperTransport, which in it's current v2.0 iteration provides 24 GB/s peak bandwidth and 3 links which effectively replace the front side bus in addition to offering a high speed link between CPUs in a multiprocessor system. If that wasn't enough, along with the rest of the HyperTransport Consortium, which includes industry juggernauts Apple, Sun, Transmeta, and Broadcom, AMD has announced the imminent release of HyperTransport 3.0, which will increase the peak bandwidth to a whopping 41.6 GB/s.
It is just this glut of bandwidth which has lead to a resurgence of interest in an area which many thought was dead and gone – co-processors. Of course, most of us use co-processors every day, as that expensive video card you shelled out for is basically a graphics co-processor, but as technology improved, the floating point and math co-processors which souped up the high end PCs of the late 80s had their functions integrated into the CPU and disappeared.
However, with many corporations looking for ways to reduce the cost of running complex algorithms, a HyperTransport-enabled co-processor may provide the answer they're looking for. According to Jeff Jussel, Vice President of Marketing for Celoxia Limited, a company which develops software for FPGA co-processors, “Banks, for example, are buying racks and racks of CPUs to run financial analysis algorithms, such as the Black Scholes algorithms. They are running out of space, and using too much power for cooling. They can offload their key algorithms to a FPGA co-processor, which consumes much less power, and see 100x improvements in performance.”
Of course, the performance increases depend greatly on the task at hand. The only current co-processor on the market which slots directly into a Socket 940 Opteron motherboard and uses the HyperTransport Bus is the DRC Coprocessor Module. Based on the Xilinx Virtex-4 family of FPGAs, it promises a 300x improvement over software-only solutions, as well as a 100x improvement in integer applications, 10x in single-precision floating-point applications, and a “mere” 5x acceleration of double-precision floating point calculations.