GPGPU: Floating-Point Performance
Single-Precision, Good. Double-Precision, Bad.
Nvidia's mainstream Kepler-based GPUs offered double-precision compute performance that was 1/24 as fast as its FP32 math. Maxwell is ever worse at 1/32. Of course, that's purely theoretical until we double-check it with real-world benchmarks.
The Folding@Home benchmark is particularly good for comparing graphics cards under OpenCL. We had to do without CUDA-based numbers this time around because the Maxwell-based card wasn't properly recognized. This is simply something we'll have to put together later.
How big is the difference between single- and double-precision, really? Our benchmark results indicate a 8:1 ratio between them on Nvidia's GeForce GTX 750 Ti (Maxwell). That's quite a bit weaker than the GeForce GTX 760's (Kepler) 4:1 in this metric (Explicit Solvent).
But GM107's comparatively strong single-precision performance is what sticks out; it's able to compete with much more potent graphics cards. By the time you get to its double-precision numbers, compute throughput ends up just below where we would have expected it.
Single-Precision Benchmarks (SP)
Double-Precision Benchmarks (DP)
We're not sure what to think about GM107's increasingly hobbled FP64 capabilities. You can either say double-precision performance is really bad, or the single-precision numbers are really good. Regardless, at the end of the day, artificial limitations meant to prevent cheap desktop cards from being viable workstation parts are no less irritating.