680LT memory performance...is it strap related?

[Tony]

For a while now it has been obvious to me that many board manufacturers have become fixated with the FSB their motherboards will do during over clocking. It seems that in the ever increasing search for 550>600fsb they will sacrifice anything to get the higher numbers in reviews.

Now many end users have bitten into this marketing hype, believing that if a board does 550+FSB then it must be faster than one that only does around 500 or lower...the simple fact is this is NOT true.

In previous articles I was able to show how chipset internal latency has a profound effect on overall speed, infact if you tweak a board correctly you can actually drop 100MHZ or so on the CPU and still end up with a faster system overall...it all revolves around the speed the memory controller handles the memory.

Now during tests of the i975 and i965 chipsets from Intel we see that forcing a lower FSB strap on the chipset than what the CPU supports really speeds things up within the chipset and is much like running 1T on the memory. The FASTEST way to run a 975/965 board such as the badaxe2 is to set 800 strap, 533 ram ratio and push the board upto around 1333fsb. Now this limits the max clock of a low multiplier CPU but if you are lucky enough to own an XE 6800 or Q6700 you can just opt for a higher multiplier and push the CPU speed to where you need it. 11x1333(333) gives you around 3.66GHZ and is a nice easy over clock, the best part though is it performs much like 3.9GHZ on the 1067 strap with memory running DDR1000. Recent benchmarks run by a friend here showed phase change cooled Q6700 running 400MHZ lower than an LN2 cooled Q6700 was faster purely due to the increase in memory performance thru the tweaked chipset performance settings, all this with memory running around DDR888.


Moving onto the NVIDIA 680i and LT chipsets we see similar trends, although they are quite well hidden within the nature of the async/sync nature of the memory over clocking settings on the board.
During initial tests of DFI's new 680LT based motherboard I was surprised to see BSEL options in bios, some investigation unrevealed the following.

BSEL= 267fsb
BSEL1=133fsb
BSEL2=200fsb
BSEL3=350fsb
BSEL4=333fsb
BSEL5=333fsb
BSEL6=350fsb
BSEL7=350fsb

Just by choosing a different BSEL option you either force an over clock or a down clock on the CPU...what had me puzzled though was why some of the options had the same effect. After some added research the following data was compiled.

BSEL 0 is infact for 1067fsb use, BSEL4 is the default for 1333fsb use.

This leave us with BSEL 3, 5 6 7 which cover a wide range of FSB's from X to Y dependent on which CPU you are using.
BSEL 7 generally is for 1800+fsb and higher with 3 5 and 6 covering the ranges from 1333 upto 1800+.

Tests were run choosing 1700fsb and sync mode for ddr850 and just the BSEL option were changed...all we were looking for was a change in memory performance...none was seen.

So moving on 680 does have strap changes so it was decided to map them out, basic 5-5-5-15 timings were chosen and the sub timings fixed so that the board would always boot with fixed settings, 800 to 1000DDR sync with CPU fsb was the chosen range.



As you see from the graph 2 FSB holes were encountered...836DDR to 842DDR and 883DDR, the board would also not boot after 986DDR, even though the ram has been tested to well over 1200MHZ in different configurations.

So what does the graph show us, well to start the memory performance discounting the 2 holes I encountered is pretty linear, although there were to bumps worth looking closer at.

DDR921 to 927 (460 to 463FSB) saw a jump of over 100MB/s which then settled back to being fully linear at DDR930, and the second bump from 971DDR onwards (485fsb). What should be noted is this second bump gave huge performance gains although the board would not boot to windows,to put the data into context just moving from DDR968(484fsb) to DDR974(487fsb) shows a 600MB/s gain in memory performance which can only be associated to some latency change in the memory controller.

So, what can we draw as a conclusion from these first test....sync mode is pretty damn fast 500fsb is quite hard to do although some will hit it and above... although how stable they will be is any ones guess.
I would try and clock as close to 490FSB in sync mode which puts the ram around DDR980, it will be tough to get it stable though but the performance returns would be huge.

Next phase is 1T tests and some windows test to see if 2T is faster than 1T.