It seems to be quite random however I have noticed on the single thread test it spikes to around 40mbps then falls back to normal levels before going up dramatically on 2nd test.
However like you said some of them are normal
Also surprised only download is affected
Browser counters getting corrupted usually produces massively higher figures e.g. 2's complement
The graph when poor is low on samples, suggests may CPU performance or browser only periodically releasing data to the scripts. Its one of those given a week with the device and connection can probably characterise what is happening.