Methodology: how we measure power draw
The 12th generation Intel Core processors – Alder Lake is now out and we already have the full-fledged tests for you. These processors are significantly different from the previous ones in many aspects and many things are used “for the first time”. Among them are DDR5 memory support, PCI Express 5.0, 7 nm manufacturing process or hybrid concept of small and large cores. It’s time for a detailed analysis!
Methodology: how we measure power draw
Measuring CPU power consumption is relatively simple, much easier than with graphics cards. All power goes through one or two EPS cables. We also use two to increase the cross-section, which is suitable for high performance AMD processors up to sTR(X)4 or for Intel HEDT, and in fact almost for mainstream processors as well. We have Prova 15 current probes to measure current directly on the wires. This is a much more accurate and reliable way of measuring than relying on internal sensors.
The only limitation of our current probes may be when testing the most powerful processors. These already exceed the maximum range of 30 A, at which high accuracy is guaranteed. For most processors, the range is optimal (even for measuring a lower load, when the probes can be switched to a lower and more accurate range of 4 A), but we will test models with power consumption over 360 W on our own device, a prototype of which we have already built. Its measuring range will no longer be limiting, but for the time being we will be using the Prova probes in the near future.
The probes are properly set to zero and connected to a UNI-T UT71E multimeter before each measurement. It records samples of current values during the tests via the IR-USB interface and writes them in a table at one-second intervals. We can then create bar graphs with power consumption patterns. But we always write average values in bar graphs. Measurements take place in various load modes. The lowest represents an idle Windows 10 desktop. This measurement takes place on a system that had been idle for quite some time.
Audio encoding (FLAC) represents a higher load, but processors use only one core or one thread for this. Higher loads, where more cores are involved, are games. We test power consumption in F1 2020, Shadow of the Tomb Raider and Total War Saga: Troy in 1920 × 1080 px. In this resolution, the power consumption is usually the highest or at least similar to that in lower or higher resolutions, where in most cases the CPU power draw rather decreases due to its lower utilization.
Power draw limits are disabled for both Intel and AMD processors, unlocked to the PL2/PPT level. As is the case with most motherboards, this is also set in the default settings. This means that the “Tau” timeout after 56 seconds does not reduce power draw and frequencies even under higher load, and performance is stable. We considered whether or not to accept the more economical settings. In the end, we won’t, on the grounds that the vast majority of users don’t do it either and thus the results and comparisons would be rather uninteresting. The solution would indeed be to test with and without power limit, but this is impossible from a time point of view in the context of processor tests. However, we won’t ignore this issue and it will be getting space in motherboard tests where it makes more sense to us.
We always use motherboards with extremely robust, efficient VRM, so that the losses on MOSFETs distort the measured results as little as possible and the test setups are powered by a high-end 1200 W BeQuiet! Dark Power Pro 12 power supply. It is strong enough to supply every processor, even with a fully loaded GeForce RTX 3080, and at the same time achieves above-standard efficiency even at lower load. For a complete overview of test setup components, see Chapter 5 of this article.
- Contents
- Intel Core i9-12900K in detail
- Methodology: performance tests
- Methodology: how we measure power draw
- Methodology: temperature and clock speed tests
- Test setup
- 3DMark
- Assassin’s Creed: Valhalla
- Borderlands 3
- Counter-Strike: GO
- Cyberpunk 2077
- DOOM Eternal
- F1 2020
- Metro Exodus
- Microsoft Flight Simulator
- Shadow of the Tomb Raider
- Total War Saga: Troy
- Overall gaming performance
- Gaming performance per euro
- PCMark
- Web performance
- 3D rendering: Cinebench, Blender, ...
- Video 1/2: Adobe Premiere Pro
- Video 1/2: DaVinci Resolve Studio
- Graphic effects: Adobe After Effects
- Video encoding
- Audio encoding
- Broadcasting (OBS a Xsplit)
- Photos 1/2: Adobe Photoshop a Lightroom
- Photos 2/2: Affinity Photo, Topaz Labs AI apps, ZPS X, ...
- (De)compression
- (De)cryption
- Numerical computing
- Simulations
- Memory and cache tests
- Processor power draw curve
- Average processor power draw
- Performance per watt
- Achieved CPU clock speed
- CPU temperature
- Conclusion
There is a small error, the i9-12900k is manufactured in 10 nm (intel 7).
We decided to treat is as a 7nm process, since it more or less reflects reality, and Intel more or less wants it to be treated like that. The “Intel 7” does not actually say “nm” anywhere, but the follow up processes will actually be marketed like that. The 20A and 18A processes actually refer to Angstroms, which are non-SI units meaning 0.1nm. So those processes are officially 2nm and 1.8nm when it comes to Intel messaging.
But when we put that aside, in my opinion it is more useful than not to use “7nm” for the Alder Lake chips. Intel’s 10nm process was in fact roughly comparable to TSMC’s N7 (7nm) process, once Intel got rid of the problems (in the SuperFin or Enhanced SuperFin versions). It is not equal parity in everything, but these technologies can be considered competitive. Calling it 10nm makes it look the process is worse than it is if people assume it is similar to 10nm Samsung or TSMC process node.
We would basically have to keep reminding people that “Intel’s 10nm is on the level of TSMC’s 7nm”. And it’s not practical to to keep repeating that every time. Ideally this should have been done earlier so that the products were not referred to as 10nm for 2-3 years in press.
In isolation, the process renaming would not be a good thing, but in a way, Intel only did what TSMC and Samsung did before them. Their processes could be said to be renamed already, so Intel just caught up with that marketing. (Not saying they are at the exact same tech level, but it’s in the ballpark for better or worse).