When analyzing the performance of your GPU in computational science, it’s very important to know that there are two levels of floating point precision: single (usually abbreviated as FP32, for using 32 bits) and double (FP64, using 64 bits). Double precision is well… more precise. Without going into too much details, consider a straightforward example:
The single precision value of Pi is 3.1415927 (7 decimal places)
The double precision value of Pi is 3.1415926535897930 (16 decimal places)
There are BOINC projects where 7 decimal places just aren’t enough for accurate computations. For example, the MilkyWay@home project, which is creating a highly accurate three dimensional model of the Milky Way galaxy, is using a lot of exponential functions in its math and all tiny errors produced by constant rounding to 7 decimal places get exponentiated by an exponential function too, delivering results which are simply not accurate enough in the end. Consequently, if you want to compute for MilkyWay@home, double precision is mandatory. And that’s where the things get tricky, because newer GPUs have actually very low FP64 performance. Since FP32 is more than adequate for gaming, AMD and Nvidia have in their wisdom decided that FP64 is for “professionals” only, so you are welcome to buy their professional graphic cards which cost much more than regular, consumer cards.
But double precision wasn’t always for professionals only. Back in early 2012. when AMD released their HD7000 consumer series, they were generous enough to equip their flagship GPU, HD7970 Tahiti, with 1.1 TFLOPS of FP64 performance, at a price of $549. In February 2013. Nvidia released their GeForce GTX Titan with 1.5 TFLOPS of FP64 performance at a price of $999 and suddenly high-end double precision computing became quite expensive. AMD followed suit and halved FP64 performance on their next line of consumer cards (Rx 300 series), relegating such niceties to expensive professional market. Good times were over (and they didn’t last long).
Not surprisingly, due to its high FP64 output, HD7970 is still an admirable BOINC performer today, very well suited for BOINC projects which require double precision computations (such as MilkyWay@home). My new machine is currently holding the third place at MilkyWay top host list which is populated almost exclusively by Tahiti cards (HD7970 is also known as R9 280X when AMD simply rebranded it in October 2013). And of course, anything that yields high performance in BOINC is also very suitable for mining Gridcoin (more BOINC work=more Gridcoins, to put it simply).