Optimum memory speed is highly dependant on the actual core speed. I'm working with blinky and chsx3 to find the optimum memory speeds for a cypress XT GPU (Hemlock, or 5970, uses the same GPU). So far I've gotten promising results from blinky's perl script. Basically, I'm brute forcing all of the normal combinations of memory and core speeds in multipels of 1 and 5 respectively. There is no set memory speed that is better than all the rest across ALL core speeds.
Chances are your miner for your 5xxx card is using phatk or phatk2 kernel, 128 worksize, vectors, and bfi_int, and hopefully yo uare using 2.1 SDK (not 2.6/2.7 which comes with the latest drivers. 2.4 and 2.5 is also suspected to be slightly slower). Here is some incomplete, but very GOOD information on what memory speeds you SHOULD be using. I have 4 5970 in 1 rig and was running 150 memory across all the cores. Using this spreadsheet, I was able to increase hashrate by 15 mhash. It's not much, but it WAS an increase. It's now doing exactly 2.7 Ghash with 910 watts, or 2.967 MHash/joule, compared to 2.685 earlier. Note that I can only recommend these speeds if you are using the above settings.