CPU speed vs RAM speed

Hi,

I hope you can help me with that question. I heard that when increasing core frequency the frequency ceiling (the max achievable ram freq) for ram decreases. Has this something to do with the IMC?
Why is this happening? I have never heard about that before and I have to admit that I do not understand this connection.

Chris

It just means that the higher the DRAM frequency, including tighter timings and a greater number of modules, the more strain there is on the IMC (Integrated Memory Controller). It can be compensated for by increasing the DRAM voltage, IMC / VTT / QPI voltage, NB / MCH voltage, loosening the timings, etc.

Ok. So that also means that less modules will reduce the strain?

And that also means that if I overclock the cpu while using lets say 2800MHz RAM at a given voltage and at stock speed I will probably run into stability issues. If I want to circumvent this I have to increase the voltage and if that does not help I have to choose higher values for the timings.

Correct?

Yes, fewer modules = less IMC strain. Unless you plan on benching w/ DDR3-2800, there's really no need for RAM rated at that high a frequency. And there are too many variables involved to give you simple yes or no answers for your 2nd and 3rd questions.

I was not aiming at precise answers. Just to get an idea.
But lets say I want to run the RAM at a certain freq but I am not stable and I cannot increase the voltage because it is not rated for higher voltages I could try to run it at higher latency instead. One would consider this approach, or?

afaiu for most applications high-freq RAMs come at the cost of higher latencies more or less equalizing the gain. So low frequency RAM with low latencies will give me a very similar performance to RAM with high freq and latency.

I've read some benchmarks that suggest at higher frequencies 1600 --> 1800 --> 2133 --> 2400 the latencies actually decrease even as the timings become looser. This makes sense as well, because the timings are expressed in terms of wait states that refer to the period for one cycle. The period for one cycle gets smaller w/increasing frequency (one clock cycle period = 1/frequency) so 14 wait states at 2133 Mhz has less latency than 10 wait states at 1333Mhz. 47ns * 13 = 658 ns vs. 75ns * 10 = 750ns

All depends from IMC. On haswell you can use memory up to ~2800 with any CPU clock and 2 or 4 memory sticks won't make big difference.
Higher clock = lower latency but higher clock = more relaxed timings = higher latency. So everything has to be balanceed for optimal performance.
Performance also depends from IMC. New platforms are scalling good up to ~2133. Above that performance gain is lower than expected so optimal memory for latest AMD and Intel chips is ~1866-2133. If you are overclocking then it can be 2400 as higher CPU clock = higher cache/IMC performance so also higher theoretical max memory bandwidth.