From: Jun Zhang (coolrainbow_at_yahoo.cn)
Date: Mon Mar 21 2011 - 19:45:10 CDT
Although I don't know the exact answer, however, I guess it may due to the float precision in CUDA? To confirm the validity, I think you should do a benchmark, I mean, perform a calculation on CPU-only and GPU-CPU version, with the same system and the same initial velocity, and the same other parameters.
> In the NAMD 2.7 Release Notes, under the CUDA GPU
> Acceleration, it is noted that forces evaluated on the GPU
> differs slightly from the forces calculated on the CPU-only
> runs. Is there a certain type of calculation that this
> will affect? Also, if I want to use GPU version, what
> kind of tests should I perform to make sure my calculation
> is valid?
> Best regard,
This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:19:57 CST