Table 4.

Average performance of 1000 queries against 1 million targets

#bits	Method	#Tanimotos (M)	chemfp 1.5			chemfp 3.3
			Avg. time (ms)	TTanimoto (ns)	Bandwidth (GiB/s)	Avg. time (ms)	TTanimoto (ns)	Bandwidth (GiB/s)
166	k = 1	91.8	0.25	2.68	8.34	0.19	2.08	10.7
166	k = 1000	588	2.20	3.74	5.97	1.85	3.15	7.10
166	T = 0.70	688	1.72	2.50	8.93	1.42	2.07	10.8
881	k = 1	146	1.50	10.3	10.2	1.22	8.35	12.5
881	k = 1000	485	5.64	11.6	8.97	4.73	9.75	10.7
881	T = 0.70	554	5.70	10.3	10.2	4.70	8.47	12.3
1021	k = 1	113	1.30	11.5	10.4	0.86	7.56	15.8
1021	k = 1000	743	9.25	12.5	9.58	6.25	8.41	14.2
1021	T = 0.70	489	5.51	11.3	10.6	3.64	7.45	16.0
2048	k = 1	356	7.76	21.8	11.0	5.29	14.8	16.1
2048	k = 1000	939	21.2	22.6	10.6	14.6	15.5	15.4
2048	T = 0.40	920	19.9	21.6	11.1	13.6	14.8	16.1

The timings use three different search methods to search the four different fingerprint types from the chemfp benchmark data set. The total number of Tanimoto evaluations is less than 1 billion because of BitBound pruning. T_Tanimoto is the average time per Tanimoto evaluation, including storing the hits. The effective read bandwidth is calculated as #Tanimotos * storage_size (24, 112, 128, and 256 bytes respectively)/T_Tanimoto. Note that while shorter fingerprints are faster and more compact, longer fingerprints tend to have better scientific usefulness