Those graphs should really start at 0, the improvements (as good as they are) are needlessly exaggerated.
They are deceptive because they exaggerate features of the data in order to improve perception of configuration changes. If every graph started at zero, many datasets would look like flat lines... not a great user experience when graphing.