Sometimes modern technology is fun, too. Especially when it involves GPGPU video cards with 128 processor cores, which can be used for parallel-computing tasks. Even more so, when modern systems support up to three of them. The combination of an interesting course on parallel computer architectures (including nVidia’s CUDA) and the availability of good deals on eBay for two additional video cards of the same make and model as the one I already had turns out to be the perfect recipe for getting started in homebrew supercomputing.
So far, I have them running the SETI@home BOINC client. At first, there was some kind of configuration problem — the work units hardly seemed to progress at all (much slower than the CPU version of the code). Since the initial estimated time for the CUDA version was much less, I figured something was wrong.
A quick Google search came up with a possible solution — disable SLI (gotta remember to switch it back on for Oblivion and Flight Sim), add “dummy plugs” to the two secondary GPUs, and extend the desktop — with Aero disabled — across the two dummy monitors.
It worked — and now the three GPUs are each crunching SETI work units roughly 20x faster than the CPU version of the code. (Overall, the relative system throughput just went from ~8 to ~70 or so.) I’m starting to see what nVidia means about the benefits of manycore computing, on problems like this.
