TABLE III.

Training performance (ms/step) for water, Cu, HEA, OC2M, dipeptides, and SPICE systems. “FP64” means double floating precision, “FP32” means single floating precision, and “FP64c” and “FP32c” mean the compressed training¹⁰⁹ for double and single floating precision, respectively. “EPYC” performed on 128 AMD EPYC 7742 cores, “3080 Ti” performed on an NVIDIA GeForce RTX 3080 Ti card, “V100” performed on an NVIDIA Tesla V100 card, “A100” performed on an NVIDIA Tesla A100 card, and “MI250” performed on an AMD Instinct MI250 Graphics Compute Die (GCD).

		loc_frame		se_e2_a				se_e2_a+se_e2_r				se_e2_a+se_e3				se_atten
System	Hardware	FP64	FP32	FP64	FP32	FP64c	FP32c	FP64	FP32	FP64c	FP32c	FP64	FP32	FP64c	FP32c	FP64	FP32
Water	EPYC	14.7	9.20	97.3	45.0	28.4	16.2	63.7	32.5	29.9	15.4	141	85.2	34.0	20.6	1210	383
	3080 Ti	7.00	4.80	24.6	10.3	9.70	6.40	26.3	11.6	12.0	8.20	52.8	17.2	16.3	6.80	199	26.9
	V100	7.90	8.50	11.1	8.20	5.90	4.80	13.6	10.9	6.90	6.40	23.5	14.0	8.60	7.30	69.6	31.7
	A100	10.7	10.0	8.20	9.30	4.90	5.70	14.5	10.8	7.80	6.30	24.5	12.0	7.50	7.20	30.8	21.2
	MI250	11.7	10.9	20.3	13.1	7.70	7.00	27.3	19.7	11.5	10.9	278	27.7	12.8	11.2	125	31.7
Cu	EPYC	4.90	3.30	33.7	12.8	8.00	5.40	19.9	10.0	10.5	5.30	45.5	24.2	9.10	6.50	226	89.1
	3080 Ti	3.20	2.20	6.50	5.10	4.60	3.90	8.70	6.30	5.90	3.40	11.8	4.80	7.20	5.70	36.8	8.80
	V100	3.20	3.80	4.20	4.80	3.20	3.70	6.50	5.30	5.50	4.10	7.90	5.60	6.00	5.80	15.6	11.9
	A100	4.00	3.90	3.80	3.70	3.10	3.00	5.40	5.30	4.10	4.10	8.00	5.60	4.80	4.60	11.6	11.2
	MI250	4.80	4.90	6.90	6.40	5.10	5.00	9.10	9.40	7.40	7.00	49.9	10.1	8.00	7.30	23.6	18.6
HEA	EPYC	⋯	⋯	53.4	30.5	19.4	12.2	52.3	29.3	27.7	16.7	83.7	51.1	26.6	15.7	159	60.1
	3080 Ti	⋯	⋯	38.4	25.2	11.2	9.10	71.4	41.8	16.3	12.7	93.6	41.0	19.7	15.0	35.9	9.10
	V100	⋯	⋯	33.2	29.8	11.8	11.1	63.2	47.4	17.5	16.5	65.5	49.6	27.4	18.7	15.6	11.9
	A100	⋯	⋯	30.5	28.6	10.9	10.4	51.6	67.4	16.9	21.2	61.7	52.9	18.6	18.8	11.7	11.5
	MI250	⋯	⋯	48.8	42.7	18.5	18.0	72.3	69.3	28.7	27.3	134	88.4	32.7	32.3	21.6	19.5
OC2M	EPYC	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	2070	625
	3080 Ti	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	352	46.0
	V100	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	120	52.8
	A100	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	51.4	30.9
	MI250	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	171	55.7
Dipeptides	EPYC	⋯	⋯	49.7	30.5	21.2	19.4	52.0	35.3	30.1	21.2	89.5	61.1	35.0	21.2	214	91.5
	3080 Ti	⋯	⋯	54.8	39.5	17.3	11.3	90.0	64.3	19.0	15.3	131	67.7	25.4	19.2	26.1	12.0
	V100	⋯	⋯	54.1	52.6	14.8	14.8	88.0	84.3	20.5	21.7	96.2	103	30.1	30.8	14.3	10.6
	A100	⋯	⋯	50.2	50.8	14.3	14.3	89.0	75.9	20.7	19.9	91.1	82.7	26.6	26.7	13.2	11.1
	MI250	⋯	⋯	66.2	67.8	23.1	22.9	117	112	35.0	32.4	155	129	45.9	44.9	19.6	16.8
SPICE	EPYC	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	244	98.0
	3080 Ti	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	35.4	15.3
	V100	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	17.3	15.9
	A100	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	11.9	12.2
	MI250	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯	29.0	24.1