OK - I can't be particularly helpful here, at least from a firmly established hardware oriented basis - 'cause, as usual, it appears to be top-effing-secret! To start out, I spent nearly an hour casting around to try to get some documentation; searching AMD tech doc database turned up exactly one item for the 790FX chipset:
http://support.amd.com/us/ChipsetMotherboard_TechDocs/A...
Read it and weep (literally!)! And then people criticize
me for criticizing
AMD and their lack of publicly available documentation, and likening the difference between Intel docs and AMD docs to, sort of, microns and light years!
Interleaving is, generally speaking, a
good thing, and you want it happening whenever possible; I have been told that this idea has led me to a mistake in setting up AMD CPU/chipset combos with selectable DCTs Mode; I automatically assumed that 'ganged' (dual-channel) would be faster, due to interleaving - but, I have been told (but, unfortunately,
not shown any numbers) that this is not
always so... I would leave it enabled for both bank and channel... For a bit more technical treatment, you could give this a read:
http://www.realworldtech.com/page.cfm?ArticleID=RWT1104...
It's a bit dated, but will give you the general idea...
As for DQS training, I found an old, old (had an AGP bus!) AMD doc that described a precursor to this, and, if they are at all like Intel, once they latch onto a design feature, my guess is they'll stick with it until it becomes totally obsolete - simply adapting the technology to faster and faster operation... The chipset contains programmable delay lines, that can be physically 'adjusted' to match RAM timing elements; these PDLs can be trained, to 'wobble', and observe the results, in order to more closely match the actual physical layout of the DIMMs, and the 'meander' of connection lines to them... The chipset can then be programmed to periodically ( like, every ten to the seventh, eighth, or ninth clock cycles) 're-train' their circuitry, but I could not figure out, from the available documentation, why this would be desirable; I also have to assume this 'disables' the memory circuitry for the amount of time it takes to 'wobble and home in on' these settings... I can't imagine that temperature would 'shift' the underlying impedances (loads) of the RAM access circuitry enough to make this necessary, and I don't know why else you'd do it... Short of trying it both ways, and running some 'synthetic' memory benchmarks at each setting, my advice here would be to disable...
CKE power down, I believe, will depend on which 'sleep state' you've enabled; if you're using an S1 sleep (power-on suspend), leave it disabled; if you intend to S3 (suspend to RAM), you probably want it enabled - again, best thing is to test - whether, and how quickly, you can 'wake up' a 'sleeping' machine; if you're like me, and leave all the 'green crap' turned off - won't matter!
Re 'tri-stating', the underlying idea is explained fairly clearly here:
http://www.cs.umd.edu/class/sum2003/cmsc311/Notes/CompO...
What it's doing on this board is anybody's guess - the only place I've ever heard of tri-stating a memory bus is on multi-processor server board, and it's pretty much been superceded by having each processor 'hooked up' to its own memory banking, and then 'hyper-transporting' blocks back and forth... I imagine this is in the top-secret, eyes-only, need-to-know basis, NDA-required, "once you read this, we have to sew your lips shut", AMD AGESA processor start-up documentation - I used to know somebody who had a copy, but I think he's in a prison camp on Titan!