Table 2.

Average time (in second) used by CPU and GPU solvers for eight selected test proteins and representative solvers. The CG and Jacobi-preconditioned CG on GPU were carried out with CUSP library and DIA matrix format, and the SA-AMG-preconditioned CG on GPU was carried out with CUSP library and COO matrix format, while the ICCG solver on GPU was implemented with cuSPARSE library and CSR matrix format. The timing scheme for each solver include all execution time of the core routine code, i.e. time elapsed on device (GPU) and on host (CPU) and on transferring data between the device and the host. Both 10⁻³ and 10⁻⁶ criteria were used for comparison.

	Protein	Ngrid	CPU			GPU
			CG	ICCG	CG	ICCG	Jacobi-CG	SA-AMG-CG
Convergence 10−3	1pmc	739431	1.81	0.21	0.32	24.65	0.26	0.57
	1e01	1010510	3.36	0.33	0.44	37.31	0.26	0.68
	1ghc	1244220	5.02	0.46	0.54	51.33	0.31	0.79
	1f53	1466600	4.73	0.32	0.53	42.48	0.30	0.89
	1e0a	1651190	4.47	0.47	0.52	60.57	0.32	0.97
	1ev0	1912380	5.48	0.79	0.61	91.38	0.41	1.06
	1dz7	2160050	7.62	0.86	0.72	106.19	0.43	1.15
	1ap0	2603130	6.75	1.22	0.67	144.31	0.51	1.34
Convergence 10−6	1pmc	739431	4.39	0.67	0.52	61.99	0.32	0.60
	1e01	1010510	7.28	0.96	0.69	84.98	0.36	0.72
	1ghc	1244220	9.87	1.32	0.86	118.73	0.44	0.81
	1f53	1466600	12.68	1.47	1.05	129.50	0.47	0.92
	1e0a	1651190	9.75	1.8	0.85	162.12	0.50	1.01
	1ev0	1912380	13.18	2.42	1.05	219.29	0.61	1.11
	1dz7	2160050	15.72	2.79	1.20	249.95	0.65	1.21
	1ap0	2603130	18.78	3.59	1.43	325.58	0.80	1.39