In the green software community we see very often that people use the TDP of the processor as a metric to estimate how much energy the CPU will consume for a specific workload.
For instance the SDIA use-phase estimation model in it’s current form uses this metric.
We have been using this metric in our machine learning model, where it serves as a very good indicator of how much the total energy consumption of the system will be.
However lately I have stumbled over a white paper from Intel where they discuss the TDP in quite some detail and even compare it to a metric from AMD that I was not aware of: The ACP.
Later in the paper Intel then continues to compare two identical machines, where one CPU is from AMD and one from Intel which have compareable TDPs.
The Intel machine does more operations AND achieves a lower power consumption. This although the ACP from AMD would suggest that the CPU has a lower energy consumption.
This comparison is a bit tricky, as when we look deeper into how the system is actually configured we do find many modifications in the BIOS that actually make the Intel CPU more energy efficient:
So the comparisons are not really on equal terrain. However it has to be clearly noted that Intel nevertheless has the better performance per watt which is probably due to the more single workload focused architecture and higher base frequency.
It shows though that the TDP alone might be a confusing indicator if energy is your concern, as it heavily depends on how the processor architecture really is, how the workload is and also how the CPU / system is configured.
Important note: The White Paper is from 2011 and when looking at the current site from Intel regarding TDP their definition reads a bit differently: Intel TDP landing page
It is unclear if the definition really changed or just the wording is a bit different due to a different author / department.
However for both cases we can conclude that the TDP should be a proper indicator of an average maximal consumption over a longer period of time. Spikes in power consumption for short durations can however be way higher.
Also TCase must be taken into account. This value is typically not given for AMD. For Intel CPUs this concludes that your CPU might consume more when you are below TCase and less when above.
The issue outlined in the whitepaper though is that a TDP can not really used for a comparison between different processor models. The range of uncertainty where the real power consumption will lie in the end is so big that easily a scenario like in the White Paper can happen, where one CPU has a lower TDP than the other, but still shows higher average power consumption.
What however has not been not been done in the White Paper is that the comparison does not take the different architecture, base frequency and core counts into account.
Our XGBoost model for instance respects this and the estimations here look a bit different:
AMD Opteron (TDP 95W), 2.2 GHz, 2 chips, 12 cores/chip, 16GB RAM
System Peak Power estimated: 241 W
System Idle Power: 93 W
Intel Xeon X5675 (TDP 115W), 3.7 GHz, 2 chips, 6 cores/chip, 16GB RAM
System Peak Power estimated: 257 W
System Idle Power: 46 W
As said before the overestimated peak power for the Intel is due to the tuning settings on Turbo Boost, which is not reflected in the Training data always.
If anyone out there has one of these Opteron processors running and they do support RAPL we would be HIGHLY interested in some measurements to see how RAPL and the official AMD measurements correlate!
Please contact uns at [email protected]