Hi CUDA collegues
Im upgrading my card from GTX275 to 560Ti (as seems best performance/$) thoughts welcome… Im using this for numerical modelling, all in DOUBLE PRECISION
My question is: does this card (and current GeForce cards in general) do double precision or is it emulated. I was under the imporession that compute capability cards 1.3 and beyond did double but it has been suggested this is not the case. Could someone enlighten me.
Much appreciated
Al
Yes, all cards from compute capability 1.3 onwards support double precision natively.
What you’ve probably heard somewhere and what might have caused your doubts is that the 560Ti has its double precision throughput artificially limited (some say crippled) to 1/12th of the single precision throughput. The way to avoid that is to buy a Tesla or Quadro card whose double precision throughput is half the single precision one. Then again, throughput often is limited by memory bandwidth, not arithmetic throughput, and the double throughput per dollar (aka bang per buck) still is better for the “crippled” Geforce cards than for the Tesla.
EDIT: You might want to consider buying a GTX 470, GTX 480, GTX 570 or GTX 580 though, which do double precision at 1/8th of the single precision throughput.
Excellent, thanks for the info and advice. ill rethink the card to buy!
Much appreciated
Al
For info and to anyone else who reads this thread Im puting a ‘run in single precision’ option into the solver (some applications this is adequate) and getting a GTX480 which seems good value for money.
Ta
Al
Mixed precision approaches are also quite useful - put the precision just where it is needed. Like doing a matrix inverse in single precision, and then iterative refinement in double precision (of course this only works if you don’t need the larger range of exponents in double precision).