Double precision

[super gnu]

Please forgive my ignorance of this subject.. I know that rt does not make use of double precision calculations, but what with the titans and tesla etc offering this as a selling point, are there any speed gains to be had there? Or is it just a matter of mathematical accuracy? I. E. Could you do two single precision calculations in one to trade accuracy for a speedup? Apologies if this makes no sense.

[Joelaff]

Is it real double precision (64 bit float)? Or is it double the half-float they had (thus 32bit) ? I would think 64bit float would be extreme overkill for most applications (at least in a frame buffer).

[3LP]

As far as I'm aware off, RT doesn't do any double precision calculation, so it's not using it at all.
I might be wrong though.

Stan

[super gnu]

My original post does state that Rt does not use double precision. I know this. I was trying to ask (obviously not very clearly) if the double precision capability of these new cards could be utilised in some way. Not to improve accuracy (obviously not necessary) but to speed things up? I. E. Doing two single precision calculations in one.

However in retrospect this probably doesn't make much sense.

[vlado]

No, it doesn't make much sense to use double precision in this way. By the way, Maxwell cards do not support double-precision arithmetic in hardware at all - nVidia removed the extra circuitry and used the space to improve the overall performance instead.

Best regards,
Vlado

[ivaylo.ivanov]

super gnu, all GPUs are optimized for floating point arithmetics. The DP hardware is just an addition to use the existing processors for DP calculations. I.e. it's the opposite - it does 1 one double precision operation instead of two float ops. Actually GTX 780 has the same hardware, DP performance is just limited artificially since it is a "pro" feauture.

[super gnu]

thanks for the clarification. i wasnt sure if it worked like that, or as you say, the opposite.