CPU Die Flatness Tests

Interesting results – Ian Anderson

As a follow up to my tests of waterblock flatness, I decided to do a retrospective review of the surface on some older processors. The results were not what I expected to find. The question this raised led to an interesting conclusion about thermal transfer.

I have tested the following CPUs using the same technique as my previous article:

Intel

  • Pentium 1 100 MHZ, 120 MHZ, 133 MHZ
  • Pentium 2 350 MHZ
  • Celeron 2×466 MHZ, 2 x 500 MHZ

AMD

  • K62 350 MHZ
  • Duron 1000 MHZ, 1200 MHZ
  • Athlon 1200 MHZ

All processors tested were within 1/100 mm of perfectly flat to within 1mm of the edge.

This result was totally unsurprising for the last three processors – the FCPGAs (flip chip pin grid array). Even the back surface of a silicon wafer is generally finished to within 10nm of perfect. Even with the protective coating, I did not expect to read any error over the 10 mm of surface.

What I found surprising were the Celeron results. As anyone who has been reading this site as long as I have knows, there is a great deal of anecdotal evidence to suggest that lapping this era (pre-FCPGA package) of Celerons will result in a boost to cooling performance. With four samples tested as unambiguously flat, I am willing to conclude that lapping could not have made any meaningful improvement to flatness.

This could mean one of three things:

  • I could have tested all four incorrectly;
  • It could also be the case that these people did not actually see a performance gain;
  • Finally, and what I expect the case was, lapping the CPU caused some improvement to performance not related to flatness.

After I tested the processors, my first conclusion was that I had made an error. As you may recall though, the test I used was self calibrating. For each data point I would test three points on the chip – one in the center and two equidistant to either side. By averaging the two on the side and subtracting the result from the reading in the center, I would get an exact and real value for the deviation from perfect in the center. The beauty of this test is it is immune from any effect of improper mounting of the CPU to the test bed.

Were I testing the processors incorrectly, or my gauge malfunctioning, I would read exactly the same value at all three points. But when I was testing the processors I read that they were at a slight tilt to the test plane. The difference I read between the two outside points was anywhere between 5/100mm and 17/100mm; were I performing the test incorrectly, these numbers would be zero.

The next most obvious source of error stems from the fact that these reports are user submitted. As such they are entirely anecdotal, unsupported, and minimally refereed, if at all. The anecdotal results could be false. I find myself disinclined to jump to this conclusion. First, These reports appear to be from a number of independent sources. Second and more importantly, there was no way for these individuals to make a financial gain.

Were these reports a hoax, it would be an exceptionally elaborate hoax for an individual or organization to undertake with no potential reward. Given that I have no way to verify or discount these reports, for sake of argument I will simply assume these results are correct and move on to other possibilities.

This leaves me with the third possibility – there is some other effect at work here that has nothing to do with flatness. There are two things lapping can do to a CPU die: thin it, and change its shape.

Most heatsinks and waterblocks were made of aluminum in the era when these processors were popular. The CPU casing on the other hand was made of copper. As has been well established on this site, copper distributes heat more efficiently than aluminum. Thus the more copper there is the more spread out the heat will be and the less ‘work’, so to speak, the aluminum and thermal compound has to do. Thus, any thinning effect from lapping the die likely reduced performance rather than improved it.

When lapping one of these processors an interesting phenomenon happens – the copper begins to show through the plating at the corners first, then work its way in. Were the CPUs being ground against a flat surface, this would seem to indicate that the processor was concave. At the time this was the prevailing assumption.

Lapping by its very nature though is a fluid process. The paper part of sandpaper is soft. Even when backed by glass, it will compress slightly and form a concave depression under the part being lapped. This will erode the edges faster than the center, leaving a convex surface bearing resemblance to a domed stadium.

What this means is you will get a small area of good contact but a lower overall heat transfer rate. In this case, that good contact area will be directly over the die. Because this increases heat transfer efficiency in the region of greatest thermal output, the total energy transfer will go up. Thus paradoxically, by decreasing efficiency you can increase thermal transfer.

Why exactly am I analyzing the performance of a processor that is the better part decade old, you may ask? For starters, this can equally apply to any modern processor with a heat spreader. The real moral of the story though is that flat is not necessarily the best shape for a thermal transfer surface. Thermal transfer efficiency may have more to do with the cooler’s ability to mate with the areas of greatest thermal load.

Ian Anderson

Be the first to comment

Leave a Reply