This project is based on the idea of creating a small supercomputer in the same vein as the Calvin College Microwulf Project . With somewhat similar project goals in mind, your task is to design a small cluster machine with as much computational capability as possible at a cost less than or equal to $10,000 (in this case, I'm increasing the dollar limit to allow for a bit more creativity). This particular project can be completed singly or as a team of two people.
There are a number of criteria on which your design will be evaluated:
Software: it must be capable of running Linux.
Cost: it must cost less than or equal to $10,000.
Maintainability/upgradability: it must be able to be maintained such that if pieces fail, then replacements can be purchased.
Absolute performance: it should have the highest possible maximum performance (we cannot really determine its sustained performance until we build it). SIMD and MIMD considerations are acceptable - i.e. you can design a machine intended for either or both styles.
Cost efficiency: try to obtain the lowest cost per GFLOPS. To further explain this point, consider that the full monetary amount does not need to be spent. As an alternative, perhaps a set of used machines of some sort could be found, enabling you to spend less and get more computing power per dollar (maintenance costs notwithstanding).
Network performance: it is important that network capability keep pace with the computational power offered by the processors. Thus, solutions for various types of higher performance networking are desirable. Parallel Gbps Ethernet, 10Gbps Ethernet, or other types of network interconnects may be desirable (but perhaps not currently inexpensive enough). One specific example: consider using a PCI express slot to increase the number of Ethernet ports on a motherboard that has onboard video.
Power efficiency: "green" solutions that get high performance per watt are encouraged. That is, power consumption should be minimized, if possible.
Suitability for high-performance computing: this reflects the idea that if it were built, we would use the computer for high-performance scientific computing activities. Thus, although high numbers of MIPS are also desirable, we're really looking for high numbers of FLOPS. That is, floating-point operations are somewhat more important than instructions per second.