Table 1.

Scaling exponents (b), coefficients of determination (R²) and completion times, calculated on the five s10up test images from 2 × 10⁶ pixels (p) to 7 × 10⁹ pixels, where time to complete t = x + ap^b. Scaling remains close to or below the theoretical optimum of O(n) and in general improves as number of active threads increases. For all but the largest image, performance gains are marginal for more than 8–10 threads.

active CPU threads	scaling t ∝ pb		time to complete (s)
	b	R2	2 MB	16 MB	128 MB	1 GB	6.5 GB
1	1.052	0.996	0.107	0.699	6.112	51.12	574.2
2	1.004	0.996	0.090	0.437	4.237	34.27	285.6
4	0.991	0.995	0.057	0.289	2.114	21.81	161.4
5	0.956	0.995	0.055	0.283	1.686	17.05	121.2
8	0.947	0.988	0.051	0.185	1.210	13.13	95.8
10	0.939	0.987	0.050	0.176	1.113	12.08	88.5
16	0.905	0.985	0.058	0.206	1.028	11.38	79.9
20	0.947	0.992	0.040	0.183	1.065	11.32	80.7
32	0.910	0.987	0.053	0.217	1.013	10.94	80.0
40	0.913	0.986	0.051	0.201	0.954	10.78	77.5