Re: Smoke Emitting FPUs
Good story there. Thanks.
I did a bonehead manoeuver, once, at the time I bought my K6/233 motherboard. Not as bad as your math Ph.D colleague's, and more cheaply recoverable, but funny.
This was back when you paid only a small premium for building up from individual components of your choosing: The price gap has widened since then. I was trying to put together a system from quality, maximum-bang-for-buck components, and so got an FIC PA-2007 mini-AT-format mainboard, K6/233, a separate huge CPU heat sink with ball-bearing fan, big tower case, PC Power & Cooling TurboCool PSU, an Adaptec AHA-2940UW, Number Nine Imagine 128 video, and 64 MB Corsair SDRAM (augmented later on, but only 64 MB initially) rated for CAS2 operation at 100MHz on the local memory bus. Carried forward from older boxen were an old Keytronics motherboard, Logitech Trackman Marble, a pair of 10k RPM IBM wide-SCSI hard drives, an old Toshiba ATAPI CD-ROM, a Toshiba floppy, and a pair of antique 3Com 3C509 ethernet cards.
I banged the machine together, flipped it on, loaded Debian, and started trying to use it. SIG11 errors started cropping up. Hmm, that's odd. Power it down, check for sources of heat problems, spaced out the cards and hard drives, made sure all fans were working, made sure there was no air blockage. Took a coping saw and made a cutout on the drive bracket, and mounted another muffin fan pointing upwards at the hard drives through it, just for an additional safety factor. Made sure nothing was jamming the CPU or power supply fans. Put the beast back together, powered up: No problems. Good. Half hour later, a consistent string of SIG11s, especially if I started doing strenuous things like kernel compiles. Damn.
Open 'er up again. Recheck for heat concentrations: None. Everything seems to be running cool. OK, am I seeing a sensitive CPU, dodgy RAM, or what? Run memtest86, no problems. Hmm. I'm being tempted, at this point, to try exchanging the CPU. But I want to try one thing out: I check out where the heat sink was mounted onto the CPU: I had a sort of heat-transferring pad between them, that came with the heat sink. So, I go down to Fry's Electronics, and buy some more of that stuff, and also a tube of heat-conducting paste, which one can use instead, in such places. Get back to my machine, take the heat sink off. Try a freshly cut piece of conductive pad. Basically the same results, except a pattern is starting to emerge: I can do the heaviest number-crunching I like within the first 20-30 minutes from power-on, without trouble. Afterwards, anything even mildly strenuous (e.g., kernel compiles) will generate Signal 11 errors. Conclusion: Something is heating up, 1/2 hour from power-on. It's not a strictly operational thing, but rather heat build-up over time.
I shut down, and take the CPU heat sink off, again. Pull the pad out, and clean all surfaces. Put the heat-conducting paste onto the bottom of the heat sink, latch it onto the CPU. Lock everything back in place, re-test. Same results. Hmm, dodgy CPU?
Shortly before I jump to that (wrong) conclusion, I pull the heat sink off, again. I put the heat sink upside-down on my desk and stare at it. Oh, fsck. What a dumbass error.
The outline of the K6 is square, as is its socket. The heat sink is also square. Unlike your Ph.D friend's FPU, the K6's socket is keyed to admit it in only one orientation, so you cannot fry it that way. So, you think, square CPU that can be inserted only the correct way; square heat sink that goes on top; lever that locks the heat sink down. Put heat sink A on chip B in socket C; what can go wrong? There are two orientations the heat sink can mount onto the CPU; 180 degree rotations of one another. Both must be right, right?
Wrong. The contact area on top of the CPU is offset towards one side, and is maybe 65% of the total square. The bottom surface of the heat sink projects downwards only on the lateral side only where the matching 65% of its contact surface can touch the CPU's contact area below it. You can verify that this has happened if/when you yank the heat sink and newly-squeezed thermal paste, because the thermal paste will be mashed from contact with the CPU. Likewise, the CPU's contact area will have some detached thermal paste smeared across it.
Only mine wasn't. On the heat sink bottom, only the paste near the centre of its contact area was mashed, with the remainder dangling into space exactly the way I'd dabbed it on. The CPU's upper surface sported only a little paste, near the middle: Most of the CPU's contact surface had been air-gapped, rather than heat-sinked. Luckily, trapped heat across most of the CPU had not turned it into silicon carne, and everything worked beautifully once I flipped the friggin' heat sink around the correct way.
The silver lining is that that machine emerged from the exercise with awesomely efficient cooling.
Rick Moen
rick@linuxmafia.com
