When I ran experiments in 2009, I found there’s a lot of noise with the FPGA’s rate due to phase errors in the encoder. Changing the encoder to 1x decoding decreased this significantly, since it always used the same edge, however it still wasn’t clean enough for us. This is equivalent to averaging over one cycle. We implemented a low pass filter and were able to get results that were good enough. Another option would be to calculate the rate from the position, which is equivalent to averaging for the sampling time.
I don’t know if the rate implementation has changed since then. Your noise seems a little higher then what I saw before, but that could easily be explained by differences in setup.
Saying that you’re running a Victor at 35% doesn’t answer Ether’s question without also defining what motor you’re using. In your chart, what are the units for “data”.