As for the heatsinks, I used my favorite and time-tested design: a sawed copper 1U heatsink for Xeons from Titan with ribs unbended like in Zalman’s coolers.
After an hour you may use the card again. Arctic Silver Thermal Adhesive is so strong that it’s real hard to tear a heatsink off the chip if you should have a mind to do that.
As for the warranty look, it’s hard to comment. I tore the heatsinks off my card and removed the remains of the glue from them. On the other hand, I read complains at forums like “I can’t return my card because I glued heatsinks to it”. Well, I guess some skill is required here, too…
Now, the most interesting part is the effect from the heatsinks. Let’s compare several states of the card:
As you see, you can’t hope for any positive effect from the heatsinks without the Vmem modification. And why should there be any effect if the chips have the room temperature (blown on by the fans)? At high Vmem, the chips are more like GDDR2 as concerns temperature. Getting them back to the room temperature or thereabouts with the help of the heatsinks, we got as many as extra 60MHz of memory frequency! To be precise, a non-stable, but allowable for testing purposes frequency was 1075MHz.
A curious fact: the chips are officially marked as Hynix AF-28, i.e. they have a specified access time of 2.8ns. Such chips are rated for work at 715MHz. Reaching 1075MHz we surpassed the specification by over 50%! As far as I know, this is an unprecedented result for memory chips, whose overclockability rigidly depends on the access time and seldom exceeds 115-120% of the rated frequency.
Of course, I’m almost sure that Hynix remarked 2.2ns chips as AF-28, but I can’t prove it: all official data say that the card carries 2.8ns chips…
Hardware modifications ended at that (later I regretted I didn’t extend them further) and the card was ready for setting records. The maximum frequencies it was stable at were 645/1075MHz – it passed 3DMark at them. But there’s also the software part of the story.