Hynix qualification has shown that write performance decreases as device get full.
See data from Hynix:
https://code.google.com/p/chrome-os-partner/issues/detail?id=54667#c51

Try localy on Kevin:

while [ $(df -k /dev/mmcblk0p1 | grep mmcblk0p1 | sed 's/ \+/ /g' | cut -d ' ' -f 4) -gt 250000 ] ; do dd of=/home/root/${i}.out if=/dev/zero bs=1M count=512 oflag=direct iflag=fullblock; i=$((i + 1)) ; sleep 120 ; df --sync -k /dev/mmcblk0p1 ; done

I massaged the output:
sed -ne '/home/s/.* \([0-9]\+\)%.*/\1/p;/copied/s/.*s, \(.*\) MB\/s.*/\1/p'  /tmp/screen-exchange | paste -d ',' - -  > perf.csv

To obtain a csv file
The graph (test was a Toshiba) does not reveal performance degradation.
Note that although stateful was full, there is space if other partition the firmware can use.
[https://docs.google.com/spreadsheets/d/1dMDhRA-mw79GZsDs8t9c8hesxaOxafZAcF_QHr4BGLQ/edit#gid=1334502935&vpid=A1]


Kevin is using a Toshiba part with cid name "500073", I could find the datasheet to match the performance.