Table 7. Scaling Across Multiple CPU Instances^a.

instances	total vCPUs	ranks × threads	PME ranks	MEM (ns/d)	EMEM	RIB (ns/d)	ERIB
1	96	48 × 2/96 × 1	0/24	127.5	1.00	10.81	1.00
2	192	48 × 4/192 × 1	12/48	158.8	0.62	20.35	0.94
4	384	48 × 8/384 × 1	12/96	201.1	0.39	37.63	0.87
8	768	384 × 2	96	182.9	0.18	59.93	0.69
16	1536	128 × 12/384 × 4	32/96	151.9	0.07	87.21	0.50
32	3072	384 × 8/768 × 4	96/192	144.3	0.04	115.49	0.33

^aGROMACS 2020 performances for MEM and RIB over multiple hpc6a instances. The third column lists the optimal decomposition into MPI ranks and OpenMP threads, and the fourth column lists the optimal number of separate PME ranks; the left entry is for MEM, the right entry for RIB if they differ.