The final say on memory (is everything we think we know wrong?)

In my time on this particular forum I've come across two different methods of determining the proper memory speed:

Halving the frequency vs. buying whatever memory you want because even DDR3-800 is fast enough to match our FSB.

However intel no longer uses a FSB on this generation of processors. What we have now is an on-board memory controller and a DMI or Direct Media Interface. Intel specifies up to 8gb of DDR3-800 2 channel memory.

This link shows that improvements in performance were realized by installing faster memory than is specced by Intel on several non-turboboosted desktop core i7s. Note that the i7-920 is configured to use 1333 and 1600 memory (which exceeds it's Intel specification of 800 or 1066 memory as it is locked to 1066MHz) and improves it's bandwidth.

This link explains how moving away from the FSB and toward an on-board memory controller got rid of the bottleneck. Note they focus mostly on triple channel memory, as you can see on page 12 dual channel mode (what we are locked to) produces roughly 1/3 less bandwidth.

So, what I'm looking for out of this thread is what speed of memory people are running on what stable overclock as well as memory benchmarks (I'll be using SiSoft Sandra lite 2011, it's free and can be downloaded from here).

I have 4gb of Alienware factory memory
I'm overclocked to 166 in the BIOs, turboboost is enabled and I do not use throttlestop.
My QPI Bus is just over 2GHz (varies from 2004-2006MHz)
JEDEC Chart:

Sandra results:
Kingston
Aggregate Memory Performance 7.42GB/s-7.53GB/s
Aggregate Memory Performance (performance vs. speed) 7.58MB/s/MHz-7.69MB/s/MHz
FSB speed 1x167MHz
Memory Bus Speed 2x501MHz
Channels 2
Multiplier x3
Maximum Memory Bus Bandwidth 15.66GB/s
Timings 6-6-6-15

Warning: Low memory bandwidth efficiency (advanced test). Generally this means a setting is incorrect or set too “safe” resulting in too low performance.
Fix: Check memory timings (reduce them) or check other chipset or memory speed settings.


I'm particularly interested in people using memory that is "faster" than Alienware factory and/or Intel specification for our processor. If I read that maximum memory bus bandwidth figure correctly...

 

Errr, the i7-920 running 1333 and 1600 memory in that link you provide is overclocked, which is why the authors were able to observe a significant increase in bandwidth. They didn't test the memory without overclocking the FSB, because that would be pointless, as has been discussed in the previous thread.

The only other interesting (but not unexpected) observation, is that using RAM with lower latencies (when all else is equal), results in a boost in performance as well. This is consistent with what everyone on this forum has been saying for a long time - to increase memory performance either (a) overclock, (b) use memory with more aggressive timings, (c) do both.

Nothing in that article suggests that if you use faster rated memory when everything else is equal, that you would get better performance out of it.

 

There is no FSB, FSB isn't even mentioned on the intel product page for our processors.

The memory controller on the 920 is locked to 1066.

Just restarted,
disabled OC
disabled Speedstep
disabled turboboost

multiplier is x9
Bus speed is 133MHz
QPI is 1600MHz
Memory is running at 399MHz
JEDEC chart is exactly the same

Sandra:
Aggregate Memory Performance: 5.8GB/s
Aggregate Memory Performance (performance vs. speed): 7.28MB/s/MHz

 

I meant to write "bus speed", rather than FSB. Regardless, that 920 is overclocked (they even give you the multiplier + clock rate in their graphs) - that's how they were able to run 1333 and 1600 memory with it at their natural speeds.

 

With a different base frequency and multiplier, this is true.

I experienced less memory performance off the overclock. It'd really be nice to see a benchmark with the Kingston CL5 memory, and then someone with a higher frequency spec memory with looser timings, it's unfortunate that no one sells 800 memory anymore.