Don't count a leading 1 as a signficant digit in the ratio.

bin/compare

 #!/usr/bin/env python
 
 import sys
 from scipy.stats import mannwhitneyu
 
 SIGNIFICANCE_THRESHOLD = 0.0001
 
 a,b = {},{}
 for (path, d) in [(sys.argv[1], a), (sys.argv[2], b)]:
     for line in open(path):
@@ skipping 18 matching lines @@
     for threshold, suffix in [(1e9, 's'), (1e6, 'ms'), (1e3, 'us'), (1e0, 'ns')]:
         if ns > threshold:
             return "%.3g%s" % (ns/threshold, suffix)
 
 maxlen = max(map(len, common))
 
 # We print only signficant changes in benchmark timing distribution.
 bonferroni = SIGNIFICANCE_THRESHOLD / len(ps)  # Adjust for the fact we've run multiple tests.
 for ratio, p, key, am, bm in ps:
     if p < bonferroni:
-        print '%*s\t%6s -> %6s\t%.2gx' % (maxlen, key, humanize(am), humanize(bm), ratio)
+        str_ratio = ('%.2gx' if ratio < 1 else '%.3gx') % ratio
+        print '%*s\t%6s -> %6s\t%s' % (maxlen, key, humanize(am), humanize(bm), str_ratio)