I have custom water cooling in my current PC, i5 2500k and for my old video card 6950, though now using a Ti 5 something i forgot, until next gen cards come out. It's fun but can be expensive.
I may go "slow" on picking WC parts, but not so much on the purchase of them. I can wait until late next year to get a 2011 board and processor. I'm leaning toward fitting everything into a HAF-922 (triple-200mm vent model), which currently houses this "test" 2700K system. so I can build the water-cooling configuration early, with parts that could anticipate the 140W+ TDP of those Haswell E's. It wouldn't make so much difference if I chose a 4790K, either. No indium solder? A delidding temptation. Whatever you could do for a 5820K, you'd want to do for a 4790K and possibly with a delidded direct-die usage. Either way, the cooling capacity needs to be somewhere between top-end AiO and an MO-RA3 without using a floor-standing external rad.
And then again . . . . I could even consider that, . . . too . . .
Meanwhile, this "old" Superclock -- "new" ACX EVGA cooler -- seems to be proving itself. Right now I'm testing with 4.6Ghz on the 2700K with 8-threaded OCCT:Linpack and a simple TRUE blue rubber-ductie (hah!) -- the "accordion" accessory. It even has minor acoustic-deadening, but needs more for this AP-30 exhaust fan. And the overall point of interest there: Linpack @ 4.6 Ghz, 1.34V severe load with droop, should show between 70 and 73C on a ducted D14 cooler -- over an ambient range between 75 and 80F. Last time I looked at the OCCT (HW)monitoring, == wait -- hold on a minute -- Yeah!! Yea-ahh, Baby!! -- 66.5C degrees, room-ambient of . . . Yea-ahh!! 76F!!
So the 6C difference reported in reviews based on average TIM and no ducting applies to a similarly ducted comparison using the same fan model, case, intake fans and thermal-control profile.
I still have to convince myself, but everyone who bought a D14 before July, 2011 missed out. It is LIKELY at this point of my tinkering and testing that the EVGA (Superclock) ACX is the better of the two. And anything better by 5C is nothing to sneeze at, either.
A single 550gm heatpipe tower beats out a D14. I may eventually try the 212+ once I have the stable clock settings -- with the same duct. I doubt it will be better than the D14, though -- and more likely worse.
UPDATE AND AFTERTHOUGHT
When I first saw the Frosty-Tech/Hardware Secrets ACX review comparison, I did not take note that the dual Noctua fans on the D14 were running at maybe 1,300 RPM each, while the ACX fan was shown at 2,200+. This may erode slightly the posted ACX gain of 6C. Instead, my own fan application -- still using the ACX fan as "pusher" -- is more like a 4 to 4.5C improvement. But I need to get better data with more tests.
And now, some advice for noobs and ASUS enthusiasts (noob or otherwise). DO NOT leave AI Suite active in the system tray while running OCCT. There will be what I want to call "collisions" in polling the thermal sensors. Either of two things will happen after an hour or two: OCCT will post an obscenely high temperature for a core, causing it to terminate because of default limits in "settings," OR . . . it will show a sensor at just above freezing. I had experienced these conflicts between AI Suite and other monitoring software before. In some rare cases, it would precipitate a BSOD. With this system and 2700K, the anomaly is more robust -- more certain. But it doesn't show up if AI Suite is completely terminated. This is the second time it happened with AI Suite active in the system tray. It just doesn't happen with OCCT running alone. Not yet. If it does, I could start to worry about "other stuff."
AND . . . ANOTHER . . .
Testing coolers for chump-change differences in load temperatures may go out of fashion, or seem less useful with custom-water. But we've never had a set of conventions for member comparisons. Some people report their highest core temperature registered in the "maximum" column of monitoring software. But the sensors are either typically out of calibration or there is some unevenness in the distribution of heat on the heat spreader. I think it is "sensor error" which is typically +/- 5C around a mean. So I believe it is both simpler and better to report a single value -- the "average of maximums." And fact is, the prevailing temperatures from captured data-tables over an entire test-run will show a lower average than that. This then begs the question as to what the REVIEWERS are using for a data reporting convention. If I WERE doing a REVIEW, I'd take all the sample data over an hour's time, load it into a spreadsheet, get the average load temperature for each core, and then average those averages.
That's the way I'D do it . . .
AND . . . YET ANOTHER . . .
Room ambient has actually increased to 77F. The retest -- after "AI SUITE" termination, assuring AVX is checked and so on for OCCT: Linpack -- is now showing a 65C Average-of-maximums after a full hour running. Should I rub my eyes? Does this actually MEAN that delta I wanted to measure is GREATER than 6C? Well -- best estimation based on experience with the sig-rig -- it IS 6C degrees with this test run.
There are reasons to love AI SUITE. And there are reasons you'd almost hate it. Nothing that can't be managed, though.