Hope everyone day is well. I need your help!
I have two cards GK110 vs GM200. I kept getting a 60% difference in performance for the same job…
The job uses only 1 Gb of GPU memory to run and complete.
I have noticed a few minor differences that could be causing an issue but not a 60% difference.
Memory bandwidth and maybe framebuffer size…
Any help would be greatly appreciated!
you can look into nvprof reports. overall, Maxwell is much more efficient architecture, may be 40% for the same peak perfromance. you may find reasons of Kepler inefficiency in https://hal.inria.fr/file/index/docid/789958/filename/112_Lai.pdf
With this little information, one could only speculate wildly. What specific GPU models are we talking about? What is the execution time for each card? Does the code in question use any double-precision computation? What were the exact nvcc compiler switches used to compile the code?