The fluctuation is caused by the way hashrate is calculated. Observe the image below.

http://i.snag.gy/ijelJ.jpg

The bar with squares represents your GPU miner thread. Each square is a GPU kernel run. Within each run it calculates --cl-global-work * --cl-local-work hashes (i.e 64 * 65536 = 4194304). The yellow bar represents the the getWork thread, which fetches new mining jobs from geth or pool and also displays hashrate. Most of the time it is a sleeping thread, with a period of --farm-recheck (default 500ms). So once in 500ms, it checks the amount of hashes the miner thread has done. But the miner thread only updates its total after having completed its batch. So the red squares aren't counted. In the sketch I quickly drafted there's only one interval with 3 runs and the rest with 4, but I hope you get the idea.

You can get a more stable hashrate figure by increasing the value of --farm-recheck (i.e. 3000), but don't do this if you're using getwork (-F flag), because then you'll also only get new jobs every 3 seconds. With the stratum miner it doesn't matter, it gets work pushed towards the miners in a separate thread. Another option is lowering --cl-global-work, but that leads to more CPU overhead, which is also not what you want.