When you first came up with the "different TjMax" theory within a Quad it sounded pretty crazy but after more testing it's about the only theory that makes any sense.
I use a Kill-a-Watt meter which shows you how much power your computer is using at the wall outlet. When running the single threaded version of Prime, whatever core I run it on, power consumption is the same. When running Orthos which is the dual core version of Prime, I get the same power consumption whether I run it on core0 and core1 or on core2 and core3. With Prime that is designed for Quad cores, I can choose any 3 cores to run it on and power consumption is always the same. If power consumption is equal no matter what side of your processor you're using then you would think that heat output of each individual core should be pretty much equal.
I decided to run 4 cores of Prime and to get the core temp a little higher I turned off the heatsink fan.
I still need to crunch some numbers but this 5C difference between core0, core1 and core2, core3 remained pretty consistent even as the first pair of cores went over 80C.
To me, a 5C difference is significant and there is no logical reason why two cores would be running 5C cooler when they are running the same core voltage and MHz and running the exact same load as the other two cores. Is TjMax different between the two pairs of cores? Maybe. I'll give users a way to individually bump TjMax for each pair of cores in the next release so other users can have a look at this issue.
It's just about time to pull the heatsink and see if that changes anything.
Edit: Here's my log file with TjMax=95C for all cores and no idle correction.
http://www.fileden.com/files/2008/3/...ncorrected.txt
Here's how the log file looks on my Quad using different TjMax for each pair of cores combined with a (+)(+)(+)(0) idle calibration. Much more realistic and likely a lot closer to the real truth.
http://www.fileden.com/files/2008/3/...empLogQuad.txt
and here's a second corrected run from idle to 80C and back to idle again.
http://www.fileden.com/files/2008/3/...mpLogQuad2.txt
Bookmarks