The GA144 data sheet has been updated at GreenArrays. They now include instruction timing and some interesting power consumption info. One graph shows execution time, power consumption and energy used for a single instruction as a function of temperature. Interestingly the power consumption and energy used goes up quite dramatically with temperature, about 30% at 70C and over 140% increase at 125C!
I'm not familiar with a mechanism that would cause that. I am aware that leakage increases dramatically with temperature, but I didn't realize the dynamic current was also a strong function of temperature. I can't imagine that the authors would not have corrected for leakage since they are only measuring the dynamic power of a single node and the static power would be for the whole chip. If the static power is not subtracted out it would greatly skew the data for applications using all the processors. But even so I don't think the curve can be explained by the leakage current.
Is this temperature dependance typical? I don't recall seeing this before.