WATT You Should Know About Processor Power
Skip to content
By OnLogic·Categories: Tech Explained·Published On: May 1st, 2020·6 min read·
When you are talking about processors with high performance and impressive stats, there is a key number you want to watch out for, especially in industrial and embedded applications: the watts. What’s watts got to do with it? You’ve seen the stat in the system — 35W, 65W, 90W, etc. Here’s what you need to know. The world of processors is a lot like the world of cars. People yell for more POWER and buy the latest and greatest without thinking. More speed! More horsepower! Damn the cost! With cars that often ends up with you holding a big bill or stuck in a ditch when you end up with too much car. With processors it’s much the same: you can end up with more “power” then you need, a processor that costs an arm and a leg, and buyer’s regret.
noun: watt; plural noun: watts; symbol: W
The International System (SI) unit of power, equivalent to one joule per second, corresponding to the power in an electric circuit in which the potential difference is one volt and the current one ampere.
Processor Power vs. Power Consumption
First off, we need to understand that power is used in two senses in the computer world. Power is often talked about as the overall processing power eg the GHz. But as important is power consumption. That second meaning is expressed in watts.
More watts is not better or worse — it’s just the amount of power it takes to run the processor at full capacity. However, the higher the number, the more your electricity bill is ticking up and the more heat is being generated. Processors that perform tasks at a greater speed inherently require more power, but more and more are optimizing their energy efficiency. To return to the car analogy, it’s like engines that perform as well, on less gas.
Efficiency and Reliability are Kings of the Industrial and Manufacturing Hill
While a processing power race is happening in the consumer space, in industrial and manufacturing industries, there is a focus on reliability and efficiency. An industrial control system for a 10-year-old million dollar machine that runs on a line doesn’t need a huge amount of processing power, like you might want for a gaming computer. Instead, it needs just enough juice to run the line of control software all-day long, every day without fail, at the lowest cost possible. Anything more is a waste, anything less is a disaster.
In terms of Return on Investment (ROI), low power consumption processors on industrial motherboards reduce downtime and increase their lifecycle. Why? For the same reason Hondas run longer than Hummers.
Low watts are also the difference between money in the bank, and bills to the power company. How much? Lets take a look.
Green is the New Black
Simply saying money can be saved or that there is money you might be wasting is not enough — let’s crunch some numbers. I have selected three processors, each with vastly different power ratings. The equation to calculate the power consumption is (A/1000)*B*24*365 where A is the watts and B is the cost of power. In the United States, the average Kilowatt hour cost of about 12¢ (EIA). Though it varies by region, I’ll be using that as my basis. Let’s see how it works out.
1 – 65 W
65W is pushing the outside envelope of what might be considered “Industrial” however it is by no means the upper limits. 90W is not uncommon in consumer boards. With it’s 65W pull, this is the biggest of this group. At .065 Kilowatts per hour (kWh), it comes to about 19¢ per day and over the course of a year, that’s $68.
2 – 35W
The 35W is closer to an “Average” industrial power pull. At .035 kWh. We are looking at about 10¢ per day and $37 per year.
3 – 14W
This last one is a a very low power one using only 14W. Thats .014 kWh close to 4¢ per day and a yearly cost of about $15.
That is maximum utilization, 24 hours a day, 7 days a week, 365 days a year. Which is not uncommon usage on a manufacturing floor running triple shifts or for a digital kiosk that runs night and day or a remote data collection station monitoring seismic activity and the like. Some industrial uses can probably lower those numbers based on their usage, but since we are focused on heavy industrial and manufacturing use, we are going to go with maximum utilization,
By that reckoning, if you are running 50 systems with 65W processors, assuming a lifecycle of 5 years, that’s $17,000. If you only need 14W, that would only cost $3,750. That is $13,250 that’s being wasted, pure and simple.
Less Heat is Much Cooler
There is a fairly complicated equation that derives the exact amount of heat that is produced by processors, but the short version is that watts produce heat. This means that some sort of cooling needs to take place so the board doesn’t melt, and the more that needs to be done to remove heat, the more complex it is going to get.
Processors are often cooled with heat sinks. They are pieces of heat conductive materials like aluminum and copper, shaped to maximize the surface area, diffuse the heat over more space and dissipate it into the surrounding area. This is why many fanless systems have fins or ribs, as the cases themselves are massive heatsinks.
However, the more powerful the processor, the harder it is for passive cooling methods to keep up. At a certain point, a fanless case will no longer suffice and one that supports a fan must be used to keep it cool. While they add a fractional amount of power to the consumption equation, even a large fan is adding only about 2w (or $2.10 a year), the bigger concern is the effect on system reliability.
We’ve talked about how the fan is the Achilles Heel of the PC , and it is never more true than here. If you have a fan, you have a point of failure. If you have a fan, you need somewhere for the air to go, which requires vents that can open your system to environmental concerns such as dust and debris. If you have a fan, you need to build the case that much bigger to make room for the fan and the larger power supply.
But if you have a low power processor you no longer need a fan. Problem solved. If you would like to learn more about fanless computing, check out our whitepaper, “The Benefits of Fanless Computing“.
In the long run, picking the right processor can save you money by eliminating wasteful power consumption. Finding the right speed, price, and lifecycle are all factoring into your decision, but be sure to keep watts in mind as well. Contact the experts OnLogic to answer any questions you might have, and learn more about low power consumption options available for you.
This post was originally published on February 28, 2014. It was updated on May 1, 2020.
Get the Latest Tech Updates
Subscribe to our newsletters to get updates from OnLogic delivered straight to your inbox. News and insights from our team of experts are just a click away. Hit the button to head to our subscription page.
About the Author: OnLogic
OnLogic is a global industrial computer manufacturer that designs highly-configurable, solution-focused computers engineered for reliability for the IoT edge.
Follow OnLogic on LinkedIn
Learn the 5 ways Fanless Computers can help your business
Get the Whitepaper
View All Articles
OnLogic Industrial Computers
Discover OnLogic’s multitude of industrial computers that will help you to advance your IoT project
Learn more at OnLogic.com
OnLogic Industrial PCs: Designed to last. Built to order. Delivered in days. Visit our online store at OnLogic.com
Page load link Go to Top
Calculating Power Consumption Of The Entire System — How Much Power Does Your Graphics Card Need?
When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works.
In this table, you’ll find a small set of standard components that you can use as a general guideline for power consumption estimates. Standard CPUs use between 65 and 85 watts, while quad-core processors range from 95 to 140 watts.
Hard drives may vary greatly according to age and model; you can get by with 10 watts as an estimate, because drives rarely run simultaneously at full load. The maximum needed is 30 watts for a short time when booting the system; you should allow a safety buffer for this when estimating maximum power load capacity.
The chipset of a motherboard can be crucial, since integrated components such as sound, network, and additional controllers must be supplied with power. While Intel gets by with 20 to 30 watts overall, a larger SLI motherboard with an nForce chipset can easily require twice as much.
Swipe to scroll horizontally
|Component||Power Consumption (Watts)|
|CPU Intel Pentium 4 (Prescott) 3.2 GHz||84|
|CPU Intel C2D E2140-2220||65|
|CPU Intel C2D E6750||65|
|CPU Intel C2Q Q6600||95 or 105|
|CPU Intel C2D E7200-7300||65|
|CPU Intel C2D E8200-8600||65|
|CPU Intel C2Q Q9300-9650||95|
|CPU Intel Core i7 920||85|
|CPU Intel Core i7 940||92|
|CPU Intel Core i7 965 Extreme||100|
|CPU AMD Athlon 64 3800+ EE||62|
|CPU AMD Athlon 64 X2 4800+ EE||65|
|CPU AMD Athlon 64 X2 4800+||89|
|CPU AMD Athlon 64 X2 6000+||125|
|CPU AMD Phenom X3||95|
|CPU AMD Phenom X4 9100e-9350e||65|
|CPU AMD Phenom X4 9500-9750||95|
|CPU AMD Phenom X4 9750-9850 Black||125|
|CPU AMD Phenom X4 9950 Black||140|
|Hard Drive 2. 5″||2 to 6|
|Hard Drive 3.5″||10 to 30|
|DVD Drive||5 to 12|
|Mainboard||20 to 60|
|1 Memory Module||3|
For a standard PC, having a powerful graphics card can easily account for 50% of the total power consumption. The values in the examples are measured liberally: the graphics card test system used has a dual-core CPU (65 nm), X38 chipset, two hard drives, and two memory modules, at 85 watts.
Swipe to scroll horizontally
|Example For A Standard PC Without Graphics Card||Power Consumption|
|Motherboard, Intel Chipset||20|
|2 Memory Modules||6|
|2 Hard Drives||20|
|Drive + Burner||20|
Swipe to scroll horizontally
|Example For A Power PC Without Graphics Card||Power Consumption|
|Overclocked Quad-Core CPU||130|
|Motherboard, Nvidia Chipset||60|
|4 Memory Modules||12|
|4 Hard Drives||40|
|Driver + Burner||20|
Calculating Power Consumption Of The Entire System
Prev Page Actual Power Consumption And Current Requirements
Next Page Connectors And Adapters For Graphics
Join the experts who read Tom’s Hardware for the inside track on enthusiast PC tech news — and have for over 25 years. We’ll send breaking news and in-depth reviews of CPUs, GPUs, AI, maker hardware and more straight to your inbox.
Contact me with news and offers from other Future brandsReceive email from us on behalf of our trusted partners or sponsors
iXBT.com CPU Power Consumption Measurement Methodology
We recently announced our new iXBT Application Benchmark 2016, which includes 17 separate benchmarks. The results of these tests allow you to evaluate the performance of the system in various use cases by measuring the execution time of test tasks and comparing this time with the time these tasks were completed on the reference system. However, processor performance is only one side of the coin, and the other side is power consumption. So far, we have not measured the power consumption of processors when testing them, but we promised to develop an appropriate methodology. In this article, we will describe a prepared power consumption measurement technique implemented using a plug-in for our iXBT Application Benchmark 2016.
Software and hardware measurement unit
The plug-in uses a specialized software and hardware measurement unit developed by Alexey Kudryavtsev. The measuring unit is connected to the break in the power circuits between the computer power supply and the motherboard. Simply put, the measuring unit is connected to the 24-pin (ATX) and 8-pin (EPS12V) connectors of the power supply, and the motherboard is connected to the measuring unit using similar connectors. The measuring unit is able to measure the voltage and current on the 12 V, 5 V and 3.3 V buses of the ATX connector, as well as the supply voltage and current on the 12 V bus of the EPS12V connector used to power the processor.
The measuring unit communicates with the computer via the USB bus. This allows you to control the operation of the block and save the measured values in a file. The operation of the measuring unit is controlled from the command line.
In our methodology, the integration of the software-measuring unit with the iXBT Application Benchmark 2016 is that in each test, synchronously with the start of the test task, data collection by the measuring unit starts, and synchronously with the end of the test task, data collection stops. The measurement results are stored in a temporary file and processed by the benchmark.
The following values are stored in each test, calculated from the measured results:
- total power consumption during the test, in watts;
- Power consumed by the processor during the test, in watts.
The total power consumption includes the consumption of the 12V, 5V and 3.3V buses of the ATX connector and the 12V bus of the EPS12V connector. The total power consumption is calculated as the ratio of the total power consumption to the measurement time.
The power consumption of the processor during the test only takes into account the consumption on the 12 V bus of the EPS12V connector (this connector is used to power the processor). This power is calculated as the ratio of processor power consumption during the test run to the test run time. However, it should be borne in mind that in this case we are talking about the power consumption of the processor together with its voltage regulator. Naturally, the processor supply voltage regulator has a certain efficiency, and part of the electrical energy is consumed by itself (it is released in the form of heat on MOSFETs and other elements). Therefore, the real power consumed by the processor will always be slightly lower than the measured values, but it is not possible to measure this real value using an external measuring unit.
Peculiarities of energy consumption measurement in individual tests
As already noted, data collection by the measuring unit starts synchronously with the start of the test task and ends synchronously with the end of the test task. It would seem that everything is simple, but there are some nuances. The fact is that some tests include not one, but several tasks at once, and the result of the test is the total time to complete all these tasks. At the same time, there can be quite long pauses between individual tasks in the test.
There are five such tests in our benchmark:
- Adobe Premiere Pro CC 2015. 0.1,
- Photodex ProShow Producer 7.0.3257,
- PhaseOne Capture One Pro 8.2,
- Adobe Audition CC 2015.0,
- WinRAR 5.21.
So, in the test using the application Adobe Premiere Pro CC , the result is the total time for rendering and exporting the movie. Tested using Photodex ProShow Producer The result is the total time to create a slideshow project, including the time to upload photos and the time to export the project to a movie. The test, using the PhaseOne Capture One Pro application, exports a collection of photos to a project with previews, batch processes photos in auto-enhancement mode, and saves photos in JPEG format. The result of the test is the execution time of all three operations. In the test using application Adobe Audition CC initially downloads a six-channel (5.1) FLAC audio file. This file is then processed by applying an adaptive noise reduction filter to it, and the final step is converting to MP3 format. The result of the test is the total time for downloading an audio file, processing and converting it. The test using application WinRAR performs two separate tasks: archiving and unzipping data.
And when several separate tasks are performed in the test, it is not clear, what exactly understand by power consumption and energy consumption, since the power consumption varies in individual tasks. Of course, you can calculate the power and energy consumption averaged over all tasks in the test, but the result will resemble the average temperature in a hospital. Therefore, we decided to proceed as follows: in the case when several separate tasks are used in the test, the measurement of power and energy consumption is made only for the most energy-intensive task. So, in the test using application Adobe Premiere Pro CC The movie export task is used. In the test using the Photodex ProShow Producer application, the task of exporting a project to a movie is also used. The test using application PhaseOne Capture One Pro uses a task to batch process photos in auto-enhance mode. The test using application Adobe Audition CC uses the task of processing an audio file by applying an adaptive noise reduction filter to it. In a task using application WinRAR only the backup task is used.
Presentation of test results
Since our real-world performance measurement methodology (iXBT Application Benchmark 2016) calculates the arithmetic mean and measurement error for a confidence interval of 0.95 for each test (the number of runs of each test can be changed), when measuring power consumption a similar approach is used. In each test, not only the arithmetic mean result for the total power consumption, processor power consumption, total power consumption and processor power consumption is calculated, but also the measurement error of these values in a confidence interval of 0.95.
Measurement results are recorded in accordance with the generally accepted rules for recording results with an error, the error is recorded with one significant digit.
Restrictions on the use of the test set
Due to the fact that the test set can only be connected to a computer power supply via a 24-pin ATX connector and an 8-pin EPS12V connector and has similar connectors for connecting the motherboard, this test set can be used only when testing systems that have the appropriate connectors. Thus, our test unit cannot be used to test laptops, nettops and monoblocks with specific power connectors. In fact, we plan to use this technique for measuring power consumption only for testing processors and, possibly, motherboards. And as for motherboards, it will still be necessary to see how expedient this is. In principle, given that the power consumption and power consumption of a processor also includes the power consumption of the voltage regulator, different motherboards can get different power consumption values when using the same processor. But it is possible that the difference will turn out to be so insignificant that such a measurement will simply be meaningless. In a word, one must first accumulate experimental data, and then make a decision regarding the expediency of carrying out such measurements.
Example of measurement results
Finally, we will show an example of a test result with energy consumption measurement.
The test bench had the following configuration:
|Processor||Intel Core i7-6700K|
|Motherboard||Asus Sabertooth Z170 S 90 087|
|RAM||16 GB DDR4-2133 (2 channels)|
|Drive||SSD Seagate ST480FN0021 (480 GB)|
|Operating system||Windows 10 (64-bit)|
|Logical group of tests||Test result, seconds||Total power, W||Processor power, W|
|Work with video content, points 0. 8.36.5757||118±2||108±2||89±2|
|SVPmark 3.0.3b, points||3300±300||83±5||64±5|
|Adobe Premiere Pro CC 2015.0.1||291±2||93±2||73.8±0.4|
|Adobe After Effects CC 2015.0.1||464±4||48.4±0.3||32.6±0.3|
|Photodex ProShow Producer 7.0.3257||394±2900 87||68.7±0.3||52 ,0±0.3|
| Digital photo processing, points
|| Adobe Photoshop CC 2015.0.1
|| 67.63±0 .09
|Adobe Photoshop Lightroom 6.1.1||319.4±0.4||91.3±0.5||70.0±0.4|
|PhaseOne Capture One Pro 8.2||373±5||59±2||43± 2|
|ACDSee Pro 8.2.287||207±2||54.6±0.4||38.3±0.4|
|Vector graphics, points||182, 7±0. 3|
|Adobe Illustrator CC 2015.0.1||356.7±0.7||39.19±0.08||24.40±0.09|
|Audio processing, points||290±3 9013 9|
|Adobe Audition CC 2015.0||360±3||61.73±0.07||46.10±0.08|
|Text recognition, points||3 85±2|
|Abbyy FineReader 12 Professional||150 .1±0.4||77.5±0.3||60.0±0.3|
|Archiving and unarchiving of data, points 0139|
|WinRAR 5.21 archiving||104 ,2±0.3||69.57±0.08||51.77±0.07|
|WinRAR 5.21 unarchive||6.8±0.4|
|File operations, points|
|Copy data||70±2||29. 9±0.4||14.7±0.4|
|UltraISO Premium Edition 22.214.171.12459||27±3||22±2||7±2|
|Scientific calculations, points||289±7|
|Dassault SolidWorks 2016 SP0 with Flow Simulation||247±6||78.3±0.4||60.6±0.3|
|Integral performance result, points||266± 6|
Let’s also present the results for power in the diagram:
The diagram clearly shows that in various tests the relative difference between the total power consumption and the processor power is different: the more the processor is loaded, the smaller this difference is. For example, in the MediaCoder test, the difference between the total power and the processor power is only 17.6%, and in the UltraISO Premium Edition 9 test.6.2.3059 it reaches 66%.
We also recall that the nominal rated power of the Intel Core i7-6700K processor is 91 watts. As you can see, in the normal mode of operation of the processor, this calculated power is not exceeded in any of the tests.
So far, we have just begun our experiments with energy consumption measurement. The current version of the hardware-software implementation of power consumption measurement has its limitations and, in fact, can only be used for testing processors. However, this alone is enough to draw conclusions about the energy efficiency of processors, evaluate how much thermal power a processor cooler should remove, and also evaluate the possibility of overclocking a processor without overheating it.
Each technique has its advantages and disadvantages. In our case, the advantages include the fact that this technique using an external measuring unit has a very high measurement accuracy and, unlike software implementations for measuring various parameters, is in no way connected with sensors and monitoring controllers on the motherboard itself and in the processor. The disadvantages of this technique include the limited possibilities of its use: the measuring unit is rigidly tied to very specific power connectors on the motherboard and cannot be used if they are absent. As already noted, this measuring unit is not suitable for testing laptops, all-in-ones, as well as ready-made PCs, since if the computer is assembled in a closed case, then connecting the measuring unit to it may be an impossible task.
However, in order to test such complete solutions as laptops, monoblocks, nettops, etc., where the use of an external measuring unit is not possible, we will soon announce another plug-in for our benchmark iXBT Application Benchmark 2016, which will allow you to control processor power, its temperature and even load in each test. This will be an exclusively software solution based on the use of a special library that allows you to get programmatic access to sensors and monitoring controllers on the motherboard and in the processor. As in well-known programs, such as AIDA or HWiNFO, during the tests, the necessary monitoring sensors will be periodically polled, which will allow fixing the average value of the processor power during the test, its maximum temperature and average load. It is not always possible to trust the readings of these sensors, however, as they say, it is better than nothing. However, let’s not get ahead of ourselves. This is a topic for a separate article, where we will pay attention to all these nuances.
Power consumption of Intel processors. Why is the energy consumption higher than expected?
Intel processors consume more than expected: heat sink requirements and turbo mode
Recently, the DIY PC community has been permeated with the topic of power consumption. The latest eight-core processors from Intel have a TDP of 95 watts, however, users are seeing how they consume 150-180 watts, which is completely not up to expectations. In this article, we will explain to you why this happens and why it causes so many problems.
What is TDP (Thermal Design Power, heat dissipation requirements)
For each processor, Intel guarantees a certain operating frequency with a certain power, often referring to a certain cooler. Most people equate TDP with maximum power consumption, given that in the calculations, the thermal power of the processor, which must be dissipated, is equal to the power consumed by it. And usually TDP stands for the amount of that power.
But strictly speaking, TDP refers to the power dissipation capability of the cooler. TDP is the minimum cooler capability to guarantee the specified efficiency. Some of the power is dissipated through the socket and the motherboard, which means the cooler rating can be lower than TDP, but in most discussions, TDP and power consumption usually meant the same thing: how much power the processor consumes under load.
Within the TDP system can be installed in the firmware. If the processor used TDP as the maximum power limit, then we would see the same measurement program produce similar graphics for high power processors with multiple cores.
In recent years, Intel has used this definition of TDP. For any given processor, Intel guaranteed the operating frequency (base frequency) for a specific power — TDP. This means that a processor like the 65W Core i7-8700, running at 3.2GHz normally, and 4.7GHz in turbo mode, is guaranteed to draw up to 65W when running at 3.2GHz alone. Intel does not guarantee performance beyond the 3.2 GHz and 65W specified.
In addition to the base values, Intel also uses turbo mode. Something like the Core i7-8700 can turbo at 4.7 GHz and consume much more power than a processor running at 3.2 GHz. Turbo mode for all cores on the Core i7-8700 processor runs at 4.3 GHz — much more than the guaranteed 3.2 GHz. The situation becomes more complicated when the turbo modes do not drop down to the base frequency. That is, if the processor constantly exceeds TDP, the 65 W cooler you bought (or the one that came with the kit) will become a bottleneck. If you need more performance, you should throw out such a cooler and take something better.
However, the manufacturer does not tell you this. If there is not enough cooling for turbo modes, and the processor reaches a temperature ceiling, then most modern processors will go into power-limiting mode, reducing performance in order to stay within the specified power consumption. And as a result, a fast processor does not reach the limits of its capabilities.
So TDP means nothing? Why has this only become a problem now?
Over the past decade, the technique for using the term TDP has not changed, but processors have begun to use their energy budget in a different way. The recent introduction of six- and eight-core consumer processors above 4 GHz means that new heavily loaded processors are exceeding their advertised TDP. In the past, we’ve seen quad-core processors rated 95W only use 50W even under full load in turbo mode. And if we add cores, but do not change the TDP designation on the package, then something must change.
Secret numbers not on the packaging
Within each processor, Intel defines multiple energy levels based on capabilities and expected operating modes. However, all of these energy levels and capabilities can be tweaked at the firmware level, leaving OEMs to decide how these processors will perform in their system. As a result, the value of the power consumption of the processor in the system turns out to be a very vague indicator.
For simplicity, three important values can be monitored. Intel calls them PL1 (energy level 1), PL2 (energy level 2), and T (Tau).
PL1 — efficient uniform expected energy consumption in the long run. In fact, PL1 is usually defined as the TDP of the processor. That is, if TDP is 80W, then PL1 is 80W.
PL2 is the short-term maximum power consumption of the processor. This value is higher than PL1, and the processor enters this state under load, which allows it to use turbo modes up to the maximum value of PL2. This means that if Intel has defined multiple turbo modes for the processor, they will only work when PL2 reaches its maximum power consumption. In PL1 mode, the turbo does not work.
Tau is a temporary variable. It determines how long the processor must remain in PL2 mode before falling back to PL1. Tau is independent of processor power and temperature (it is expected that when the temperature limit is reached, a different set of ultra-low voltages and frequencies will be used, and the PL1/PL2 system will stop working).
Here are the official definitions from Intel:
Let’s take a look at the situation of high CPU load.
It starts in PL2 mode first. If the load is single threaded, we should reach the high turbo value that is indicated in the specification. Typically, the power consumption of a single core will not come close to the PL2 value of the entire chip. If we continue to stress the cores, the processor will respond by reducing the turbo frequency according to the per-core values determined by Intel. If the processor’s power consumption reaches PL2, then its frequency changes so as not to go beyond PL2.
When the system is under heavy load for a long period of time, «Tau» seconds, the firmware should jump to PL1 as the new power limit. Turbo tables are no longer applicable — they only work with PL2 mode.
If the consumption is outside PL1, then the frequency and voltage are changed so that the energy consumption remains within these limits. That is, the processor entirely reduces the frequency from the PL2 state to the PL1 state for the duration of its operation under load. This means that the temperature of the processor should decrease, and this should increase the life of the processor.
PL1 mode runs until the load is removed and the core goes idle for a certain amount of time (usually up to 5 seconds). After that, the PL2 mode can be turned on again when another large load appears.
Here are some examples of values - Intel lists several options in the specifications of various processors. For example, I took the Core i7-8700K. The following is true for this processor:
PL1 = TDP = 95 W
PL2 = TDP * 1. 25 = 118.75 W
Tau = 8 sec
In this case, the system should be able to accelerate to 119 watts for eight seconds, and then roll back to 95 watts again. This has been the case for several generations of Intel processors, and for the most part, it didn’t really matter, since the power consumption of the processor as a whole often turned out to be well below the PL1 value even under full load.
However, all the bullshit starts when motherboard manufacturers come into play, since PL1, PL2 and Tau are all configurable in the firmware. For example, in the graph above, you can remove the restrictions from PL2, and assign 165 W and 95 W.
The world of random numbers
I will mainly talk about consumer electronics. Often PL1, PL2, and Tau are carefully controlled in cooling-limited environments such as laptops or small PCs. I’m familiar with several powerful yet stylish PCs that also have PL2 set to TDP so that the processor can overclock a little, but not to the point where one or two cores are loaded beyond TDP.
However, in our CPU reviews, after the proliferation of six-core processors, we often began to see numbers much higher than PL1 or PL2, and this consumption continues for an arbitrarily long time, unless it goes beyond temperature limits. Why is this happening?
In any modern BIOS, especially those of major motherboard manufacturers, there will be settings for limiting power (short-term and long-term) and duration. In most cases, by default, the user does not know what value they are set to, because it will say Auto, which is the code for «we know what value to assign to them, don’t worry.» Manufacturers will store the values in memory and use them, but the user will only see Auto. As a result, you can assign PL2 to 4096 W and make Tau very large, for example, 65535, or -1 (infinity — depends on the BIOS version). This means that the CPU will run in turbo mode without interruption until it exceeds the temperature limits.
Why do manufacturers do this? There can be many reasons for this, although specific reasons may vary from manufacturer to manufacturer.
Firstly, this means that the user can keep the turbo mode running all the time, and each core will run in turbo mode every second. The results of performance measurements will reach the sky, in reviews or when the user is measured by indicators, everything looks great,
Secondly, products are developed for this. Intel often defines a default motherboard specification with each launch (they even had their own motherboards that they sold at retail), with a certain number of power phases, and an expected lifetime. Manufacturers can obviously implement their own options: more power phases, more powerful phases, special power supply to improve efficiency, etc. If their board can support all-core turbo all the time, then why not?
Thirdly, manufacturers of more expensive motherboards know that enthusiasts will use improved cooling systems for them. If the processor draws more than 160W and the user has a decent cooling system, then turbo mode on all cores will improve the experience of the product. Intel standards are defined for the coolers recommended by the company.
So what’s the right thing to trust, what’s the difference?
Intel sets standards for its parts. PL1, PL2, Tau, motherboard layout, firmware settings — everything has default values recommended by Intel. Some of them are public, such as those that Intel indicates in the documents, some are confidential (and Intel will not tell us about them, no matter how hard we beg). However, these are still recommended values. And as a result, motherboard manufacturers can do whatever they want. And they do.
As a result, for example, it becomes more difficult for me to test equipment because of this. Different users will want our settings to be:
1. Recommended by Intel,
2. Out of the box,
3. Turned to the maximum.
And, of course, Intel’s recommendations will give much lower results than «out of the box», and the «turned to the maximum» option speaks for itself.
It is worth noting that so far in all tests in all CPU reviews, hardware has been run on out-of-the-box settings, and not on those recommended by Intel.
To give some context to the measurement values, we used a powerful CPU and
got the following results in a 25-30 second full load test:
Intel Spec, 165W
Intel Spec, 95W
Lately, it has been noticed that some motherboard manufacturers are changing their PL1/PL2/Tau strategy and cutting Tau down to something reasonable like 30 seconds. When running speed measurements on these motherboards, users get less than usual results, although these results are closer to Intel’s specifications.
The fact is that when motherboards are set to Auto, the manufacturer usually does not disclose the exact value of this value. As a result, it is very difficult to describe the operation of such equipment. Also, these values may vary depending on the installed processor.
We usually test with out-of-the-box settings, with the exception of memory, with which we use the manufacturer’s recommended values. We believe this is the most honest way to let readers know what speed they can expect when virtually no settings have been changed. In reality, this usually means that PL2 is set to some very large value, and Tau is set to a very long value. We constantly encounter turbo mode as long as the temperature remains within the established limits.
The current situation and what we can do about it
I have long wanted to write a similar article, at least since the launch of Kaby Lake. Most processors in consumer motherboards run at unlimited PL2, and this has been considered normal for years. It was only after testing the Core i9-9900K that we began to notice something strange. In our article last week on the new Xeon E, our Supermicro motherboard literally follows Intel’s guidelines. It might seem obvious that a more commercial/server board would follow Intel’s specs, but this is the first time I’ve seen this in person. It is obvious that consumer boards according to such specifications do not work, and did not work. I’d argue that Intel’s own test results (and AMD’s Intel CPU test results) on consumer motherboards don’t match Intel’s specifications either.
So what do we do about it? I would say that Intel should put two power designations on the boxes:
- TDP peak for PL2
- TDP long term for PL1.
This is how Intel and others can explain peak consumption and base frequency.
If users want consumer motherboards to change, it will be more difficult to do so.