cuda - Estimating increase in speed when changing NVIDIA GPU model -
cuda - Estimating increase in speed when changing NVIDIA GPU model -
i developing cuda application deployed on gpu much improve mine. given gpu model, how can estimate how much faster algorithm run on it?
you're going have hard time, number of reasons:
clock rate , memory speed have weak relationship code speed, because there lot more going on under hood (e.g., thread context switching) gets improved/changed new hardware.
caches have been added new hardware (e.g., fermi) , unless model cache hit/miss rates, you'll have tough time predicting how impact speed.
floating point performance in general dependent on model (e.g.: tesla c2050 has improve performance "top of line" gtx-480).
register usage per device can alter different devices, , can impact performance; occupancy affected in many cases.
performance can improved targeting specific hardware, if algorithm perfect gpu, improve if optimize new hardware.
now, said, can create predictions if run app through 1 of profilers (such nvidia compute profiler), , @ occupancy , sm utilization. if gpu has 2 sms , 1 run on has 16 sms, see improvement, not because of that.
so, unfortunately, isn't easy create type of predictions want. if you're writing open source, post code , inquire others test newer hardware, isn't option.
cuda gpu-programming time-estimation
Comments
Post a Comment