So I spent about an hour running my straightedge over it to polish and flatten the surface further that way.  The straight edge is steel so copper shouldn't be able to hurt it. 
 
What I noticed was there was a burr around the top hole.  Everywhere that is dark in this picture is where material was removed.  Interesting you can see the outline of the die:
 

 


It's as close to perfectly flat as I think it can get.  I'll have to remount it again either later tonight or tomorrow.  I'll probably try with Liquid Ultra as well.
 
 

Scarlet-Tech
Liquid ultra is going to make a mess if you try it. Stick with the known thermal marterial to get a proper reading. Once you are ready to mount it permanently, then go with ultra. Be careful with ultra, because I am pretty sure there were complaints of it drying out and fusing to the heat sink.

Also, I found bare mounting my 4770k vs using the ihs on it, using CLLP during bare mount and using CLLP between ihs and Mx between ihs and cooler, is that the temps were the same. Hopefully you can get some better results.

I think the "maybe 400 grit isn't enough" should have been a known. I went up to 1500, and was finding streaks and minor impressions left with even 800. Go higher, make it smoother, and then give it a shot.

In a roundabout way I have the CLU because I used to direct die cool my 3770k and 4770k.  Waterblock right on the die with CLU between them.  I would be able to slam 1.6V into the chip and not get above 70C in P95. 
 
Figure this will be the last attempt before I send it back and have nateman aim to completely remove the IHS and hopefully not destroy silicon :)

The lapping was done on a grade A surface plate with a flatness tolerance of +/- 0.00005".  That is way more accurate then my home-made CNC.  Remember I don't own a HAAS, or other 100K VMC, this is a home made machine that can get  +/- 0.0005 on good days after I tram it up, and 0.001 on normal operation.   That chip lost all its rigidity with half its copper machined away.  
 
I would say there could have been warping of the material during the machining process by removing the surface tension of the top of the block.  But then I lapped it flat.   I don't think 400 grit is too course, I have done incredible things with chips @ 400 grit.  
 
The chip seems to make good contact with the block, but perhaps the block needs to be lapped.  Don't mistake a shiny waterblock as a flat one.  The surface looks polished, which uses a buffing wheel, not a flat surface.  
 
Every process you do to the chip now I think is now superficial.  The socket itself isn't machined steel, its plastic, the PCB isn't a rigid structure etc.  At least when there was more copper in the IHS, it had some rigidity (albeit copper not steel which has more rigidity, but still has to bow to a granite surface plate)

Tweaked
I think that deserves a BR.

I didn't even notice... thanks!
 
Perhaps you need a different CPU block:

 
you won't have to worry about pressure if you have the CPU block beveled to the size of the die.  Can't really deny good contact with that setup.  The CPU block itself will have some structural rigidity (that would be a first). 
 
 
 

Scarlet-Tech
I would think the hole is for pressure, since heat would build pressure over an extended period of time if the entire ihs was sealed.

Don't forget that the air pressure at the factory where it was sealed may be different, a pressure can build up even when the CPU isn't operating by simply shipping it somewhere that typically has a lower pressure (i.e. a colder environment or higher elevation).  Same deal for ball-point pens having little holes, never know where they'll be sent/used.

Bring back the old school...
 
Awesome work guys....

Yes, lapping both a block and cpu flat ensures the most surface to surface contact for heat transfer. 
 
When I said why lap, when you can machine... was just sort of a catchy title indicating that I was machining a cpu (scarey as hell being the first time).  You are correct, lapping on a granite surface plate is far more accurate then what my machine can produce (though there are machines that can duplicate lapping quality)
 
The goal for this project was to remove more mass from the CPU so that the transfer would go more quickly into the cooling device, whatever Anonymous chooses to use.

Both lapping and machining mostly serve the same goal.  Lapping reduces the thickness of thermal paste required to mate the two surfaces (it might increase total surface area a little bit, but the mounting pressures are high enough that it's not much change).  Machining reduces the thickness of IHS.  Because of this:
 



basically indicating your heat transfer is a function of the temperature differential (T_hot - T_cold) and the thickness of the material (d), you can either make your liquid/waterblock colder or decrease the total thickness to get better heat transfer.
 
At the first stab there probably wasn't much improvement because while I decreased (d) substantially, I probably significantly increased the thickness of the thermal paste layer with the surfaces not mating up very well (pressure warpage, not being flat, etc).   Because we're dealing with copper IHS and the thermal paste, the thermal conductivity has to be looked at as a weighed average between both of them.  If before d was 10, with 90% being the IHS and 10% being the thermal paste (I'm just guessing 200 "units" for thermal conductivity of copper vs 8 for paste), you'd get 90% * 200 + 10% * 8 = 180.8.  After machining down the IHS d is 3, 60% is now IHS and if 40% is paste you get 60% * 200 + 40% * 8 = 123.2.  Average thermal conductivity through everything decreased from 180.8 to 123.2 because the thermal paste got thicker and dragged the average down.  "ah ha" you might say "but the thickness decreased".  OK so 180.8 conductivity / 10 thick = 18.08.  123.2 / 3 thick = 44.3.  So we only improved thermal conductivity by a factor of ~ 2.5x even though the total thickness was reduced by over 3x.
 
Hopefully all that makes some vague sense.
 
 
Also update on this: spent 2 hours last night and 8 razor blades scraping the IHS until it's ridiculously flat.  One corner is about 3.629mm thick compared to the rest which is about 3.57mm thick, but I think its a case of the surface being true-but-not-level.  My straight edge even sounds different and seems to float over the surface when I drag it.  I'll plug it in tonight and cross my fingers to hope temperatures actually improved.

^^ we had to do it in one shot since nateman doesn't have an X99 or watercooled system to work with so there already is a fair bit of shipping parts back and forth going on.
 
The ideal state isn't really a secret though - if it could be done without risking silicon I'd have had nateman go straight to completely machining off the entire IHS down to the solder on the die.  We want to get as much cooling surface area in the same plane as the heat source (the silicon) as possible.  The silicon is only something like 70 microns thick so if nateman's machine only has z-axis tolerance of .001" then that's 25 microns right there of uncertainty.
 
LN2 you probably need some thermal buffer because it boils off really quick and probably has boundary layer issues, but with a waterblock that shouldn't be a problem so less material is better.

Update: there's a problem with thermal paste spreading.  I tried again and temps still weren't great.  I then applied thermal paste to the top with my straight edge and scraped it off leaving behind a thin film, and the hottest core temp dropped 9 degrees C.  CLU might be the way to go since its pre spread anyways.  It might be the waterblock and the cpu are combined flat enough that there's not enough central clamping pressure over the die to thin out the thermal paste enough. Or just remove the IHS completely and then all the contact pressure is directed on a smaller area.

So I think this phase of the experiment is complete.  CLU was the best result thus far, but it was still pretty much the same as stock IHS (at least on the hottest core...I wasn't really looking at averages since the hottest core is what holds everything back).  If nateman is up to it, I figure we go all in and try to completely machine / lap off the IHS.  Not that it matters - but I don't think anyone has ever completely lapped off an IHS so #first.
 
Here's some pics of last nights endeavors:
 
About to put waterblock on.  You can see I had to use small nuts and washers to keep the backplate held on.  There's a very thin scraped layer of thermal paste on it, which was the best result I could get with thermal paste.  The X method looks like it doesn't thin the paste layer far enough.

 
Here's what the thermal paste contact patch looked like after removing waterblock.  Hard to believe this was the best temperatures.

 
CLU after the waterblock was installed and removed

 
CLU again.  You can see one of the quick disconnect fittings which was why I was able to do swaps reasonably quick...waterblock stayed filled with fluid.



I bought the gold plated waterblock specifically because of CLU dissolving nickel (etching it really) and thinking "oh gold is impervious to everything".  Nope  waterblock and IHS are both stained silver now.  Alcohol, brake cleaner, and Xtreme Clean have no effect.

Awww well second place still isn't bad :)  Nice find though.  Probably the best strategy will be machining off everything but a foil-thick layer of the IHS then lapping it off.  With the machine we can also take the IHS edge closer to the PCB so there's no chance the waterblock contacts that instead of the die.
 
The question mark is I don't know how scratch tolerant silicon is.  The top of the die is the back of the wafer, so it's not like all scratches are bad, but no idea what depth of scratch is dangerous.  I remember reading specs somewhere for quality analysis about what scratch depths are acceptable but I can't find them now.  Nvidia engraves their logo and model information on their die, but they probably leave extra meat on the wafer in the backgrinding process.
 
Also just to keep everyone filled in, the 5820K is just the sacrificial part for this process.  Nateman is still holding on to my "good" 5960X that will get the same treatment once we find which treatment works.  The 5960X my machine currently is running is a worse overclocker.

acetone gets everything off ;) 

nateman_doo
acetone gets everything off ;) 

I heard hydrochloric acid, get rid of everything.

Scarlet-Tech
 

 

That would be absolutely AWESOME to have a 5960x like this!  Hmmmm, wonder how that would work with Phase.... sorry still have Hallowen's phase setup in my head from earlier.
 
Hope it works out well on your 5820, that 5960x like this would be epic.
 
And leave it to nateman to CNC a CPU....