There are two types of throttling. First there is turbo throttling where a Core i7-920 will start to drop the +1 turbo boost and it will cycle the multiplier from 21 to 20. This can happen hundreds of times a second. This happens on the Asus P6T based boards when significantly overclocking them and the core temperature reaches about 80C.
For thermal throttling, this will happen close to the official TJMax temperature. On Core 2 CPUs, it tends to happen about 2C or 3C before the official TJMax number. When the Distance to TJMax shows 2 or 3 in RealTemp, the thermal throttling bit will be triggered and in the Thermal Status area of RealTemp, it will change from OK to LOG which shows that at least one thermal throttling episode has been logged.
The actual core temperature can change very rapidly when there is a problem with a heatsink or what have you. RealTemp samples the CPU once per second so it's possible to trigger thermal throttling without RealTemp reporting 100C. When the multiplier cycles down to the default of 12, the temperature can drop back down instantly as well. There is a lot of energy flowing through a very small device so rapid temperature changes do occur.
I wish I had a few more details from Intel but I don't. The rated TJMax is contained in MSR 0x1A2. The Core i7 CPUs I've seen tend to trigger the thermal throttling bit within 1 degree of the number in this register.
If the temperature is changing rapidly, RealTemp might report 95C even though the temperature quickly went beyond that temperature before dropping back. To test you would need to increase the core temperature in a very slow and controlled manner.
If you use the TJMax that is written into the CPU, your reported core temperatures for core 0 will be extremely accurate from idle to TJMax. The sensors in the Core i7-920 are excellent. They don't have the problems that the previous 45nm Core 2 sensors had. It's my opinion that Intel sets the TJMax 5C higher on core 3. If Core 0 is set to 100C in the MSR then in almost every case, core 3 will have an actual TJMax close to 105C regardless of what the register says for that core. The two center cores might have a slightly different TJMax but they tend to be fairly close to core 0. I believe Intel does this deliberately to better control thermal throttling so all 4 cores don't reach the thermal throttling point at the exact same time. Just a theory I have and there is no documentation from Intel to support this idea. If you see enough screen shots of Prime95 running Small FFTs then you might agree with this conclusion.
12/02/09 03:12:19 PM]