Computation time of GTX 860M


I would like to ask that the computation time becomes very slow when using GTX 860M (Maxwell, compute capability 5.0) as compared with GTX 740M, the environment settings are

  1. Visual studio 2010
  2. CUDA sdk 5.5
  3. Grid size : (512,128,180)

Do anyone know the reason? In addition, I couldn’t use the Nsight to evaluate the system performance since it break down during testing. However, the original code can be compiled and run. Is there any extra installation while using Nsight? Thanks a lot!!

What is the relative performance between the GPUs? What kind of application are you running and measuring the performance for? Is it memory bound or computation bound? Integer or floating-point computation? Are the two GPUs used inside the same system, or are they in different systems, meaning the application-level performance differences could be due to non-GPU factors. Is the same CUDA version being used with both cards, and is the code compiled appropriately for each GPU’s compute capability?

Since these devices are presumably inside a mobile computer, make sure you perform all performance measurements when the computer is plugged into a wall outlet, not running off batteries. When running on batteries performance may be reduced significantly by the power management software causing the GPU to run at lower clock rates. This slowdown may be different for different GPUs and/or host systems.