Hi! Ray and Trung
Thanks for your reply！
(1)In my system, there are 808800 atoms and chain length is 1200, and I run 100 steps. (2 MPI taks)
If I don’t use GPU, the ave breakdown time : pair/bond/neigh/comm are 10.16/1.61/7.33/1.92, respectively
If I use GPU commad, the ave breakdown time : pair/bond/neigh/comm are 4.88/1.84/13.97/1.95
If I don’t use GPU but add “newton off on”,time: pair/bond/neigh/comm are 8.87/1.52/11.71/1.89
If I use GPU and add “newton off on” ,time: pair/bond/neigh/comm are 4.73/1.70/13.44/1.86
Just compare about two GPU results, there is little benefit when I add “newton off on” in input
To my knowledge, if I turn off the pair newton, the program would use same neighbor list, and there is some acceleration
But it is not enough for my problem
(2)KK package is another way, but I could not use KK in windows(Axel told there are some problem), and before our group install the gpu in our cluster, I must give a confirmed answer to my supervisor that GPU is helpful. Therefore, I buy one in my desktop and test it. And I have to learn how to install Linux in my desktop, then Lammps, but it’s not easy for me beacause of some reasons
In fact, KK package must make MPI tasks equal to GPUs, I believe this method could not use full resource of a node
Anyway, I am very happy talk about this problem with yours
Best wishes
Yongjin, Ruan