The IMC is a part of the northbridge, and the stock NB speed for Kaveri is 1.8 ghz. The OP should be able to push it to 2 ghz without trouble. Anything beyond that seems to require going sub-zero, since the only 2.4 ghz NB I've seen on Kaveri was under phase.
Don't be fooled by speed ratings into thinking the IMC will have trouble servicing fast memory. DDR3 uses many "tricks" to run at effective speeds far above its actual speed. The speed rating you see on DDR3 is actually 8 times higher than the clockspeed of the RAM itself, never mind that running in dual-channel mode effectively doubles the bandwidth again. Furthermore, the IMC has tricks of its own to handle reads from/writes to memory far in excess of what you might think the IMC can handle.
The primary performance problem from a slow IMC on an AMD chip is the bandwidth/latency of the connection between the IMC and the cores. I suppose you could saturate the IMC if you underclocked it and ran some incredibly-high-speed memory, but that's not likely to happen.
If you were seeing no real improvements from running higher memory speeds, then it may have something to do with the fact that, often times, reaching higher memory clocks comes at the expense of timings. Such an adjustment should, in theory, allow higher bandwidth at the expense of equal-to-higher latency. AMD chips are usually (read: usually) not starved for bandwidth, but rather are starved for quick memory access (AMD cache tends to be slow), and this all falls at the feet of the IMC eventually.
So, when tuning an AMD chip for performance, some decent gains can be had from running the IMC at a higher clock speed and tightening the RAM timings, even if you have to give up a little bit of memory clock speed in the process.
When it comes to Kaveri (or any other APU), the iGPU IS starved for bandwidth, but it still has to deal with the IMC behavior which often produces higher effective bandwidth at tighter timings, so you have to balance between timings and clockspeed very carefully. My guess is, the IMC is still predominantly configured to serve a general-purpose CPU (such as the four Steamroller cores of Kaveri) rather than a GPU which can more-easily make effective use of large sequential reads/writes from/to memory.
If memory serves, sequential read/writes are not terribly latency-dependant if the controller is set up to facilitate them under high-latency conditions. AMD's IMC does not appear to be thusly configured.
Facebook
Twitter