Very, very cool. I'm always happy to see more ways to utilize the tremendous performance/price of GPUs.
So, what are the pros and cons of this approach? For a given hardware and energy budget, what kinds of task would be better suited to this kind of GPU implementation? Conversely, what kind of tasks would still be faster to run on a Core i7?
I remember some time back when GPUs were first being looked at for HPC uses I suggested a client look into them, but they were quickly dismissed due to their floating point limitation (which may have changed).
But DBs don't have huge FP needs. Another place GPUs are making a big difference is video editing, utilizing CUDA or OpenCL.
GPUs have always had amazing floating point performance for single precision floats, that's basically all graphics is. The problem is that double precision support was traditionally non-existent, and has been slow until recently and it is still much slower than single precision float performance.
Double precision performance is now just half speed, which is the same as the CPU. Many applications are bandwidth bound, so running in double vs. single doesn't end up with a 2x difference anyway.
So, what are the pros and cons of this approach? For a given hardware and energy budget, what kinds of task would be better suited to this kind of GPU implementation? Conversely, what kind of tasks would still be faster to run on a Core i7?