This is the BIOS writer's guide for Athlon64. It has a section with the
timing parameters. Not all of the AMD timing parameters are a function
of the memory - some are specific to the motherboard implementation.
Those would best be left at auto. (Pg.81)
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/
26094.PDF
Here is a sample memory datasheet. There are five figures of read
and write cycles at the end of the datasheet. There is a table of
timing numbers (in nanoseconds - these have to be converted to
clocks at 5 nanoseconds apiece). Notice how some of the timing
parameters are interdependent - one timing parameter must be
greater than the sum of a couple of others. This means you cannot
arbitrarily crank down the numbers. If you do crank down the numbers,
I think the burst of data can be aborted, and the memory bandwidth
suffers. (Possibly a low Tras will do that.)
(pg.58 DDR400 3-3-3 timings, figures pg.77-81)
http://download.micron.com/pdf/datasheets/dram/ddr/256MBDDRx4x8x16.pdf
I have yet to see an authoritative description of setting timing
parameters, so I hope the above two docs will give you some hints
as to how it works.
At PC3200 (DDR400), the definition of the SPD EEPROM is covered by
the standards. That means the proper timing should already be there
without having to set it. (I.e. Corsair would have put DDR400 timings
in the SPD on the memory DIMM. For PC3500, PC3700, PC4000, this is
normally not the case, and it is possible for those sticks, the SPD
is still set to numbers appropriate for PC3200/DDR400. In those cases
it is up to the user to set the timings.) Perhaps a Windows utility
like CPUZ or Sandra, can verify that you've already got 2-2-2-6 ? If
that is the case, it is pretty hard to twist the knob much tighter
than that
If you are cranking the FSB, you'll have to back off on the timing
numbers a bit. A good place to start, is changing Trcd from 2 to 3,
as that is one way to encourage a higher clock. You should also set
Vdimm according to the manufacturer's voltage limits - overvolting
memory that cannot take it, is an expensive mistake.
Check a review site like Anandtech, for recommendations on Tras.
Wesley Fink usually runs tests to determine the range of numbers
which give good bandwidth numbers. A higher number can sometimes
be better, but it is a function of the chipset/BIOS, so someone
else's experimental results are a good source of info.
http://www.anandtech.com/mb/showdoc.aspx?i=2128&p=5
HTH,
Paul
