Quote:
Originally posted by CodeRed
@HKPolice,
I see a couple of problem with using the sandra mem banchmark as a reference.
Firstly, 10MB/sec is only a small gain and possibly due to the variantion between runs.
Secondly, the sandra mem benchmark has a very sequential memory referencing pattern. It is very localised and greatly benefits from in-page access cycles. By enabling the dynamic idle cycle counter the mem conroller is probably prematurely invalidating pages and thus inccurring a performance hit. If this is truely the case then try disabling the dynamic idle cycle counter (the mem controller will then use a static counter algorithm) and set the Idle Cycle Counter to its max value (256 clocks). This should ensure that mem pages remain valid for as long as possible. Test it and let me know how it goes.
Thirdly, sandra sucks as a benchmarking tool :D It really measures the burst speed of your memory. Try it unbuffered and I am sure the results will be different.
@the.m,
Most memory with the A64 only works at either CAS2.0 or CAS2.5/3.0. Not many sticks work with both cas2.0 and cas2.5 (XMS3500C2 may be an exception here). My kingston sticks wont run anything but cas2.0 or 1.5. CAS2.5/3.0 just freezes. There may be a combination of parameters that helps out, but I dont know what that is (yet).
@EmineM,
cpulloverclock is working on a universal tweaker, so drop him a line.