AMD and Intel have never been as competitive as they are in history. In our latest Processor Best Buy Guide, their top models differed only 3 percent in the Performance Score. Only good for us consumers, you would think, but this fierce battle has a dark side; the latest CPUs consume more and more power and are therefore getting hotter.
In this how-to, I explore the possibilities of making your processor work more economically and show you what the consequences are for performance, power consumption, and temperatures. I do this using two of the most consuming consumer processors at the moment: the AMD Ryzen 9 7950X and the Intel Core i9 13900K.
Power limits and undervolts
The power consumption of the processors mentioned is of course pure madness. In our measurements, AMD’s top model consumed more than 226W, while the Core i9 13900K far exceeded that with almost 339W. A moment of realization for those who have been involved in the hardware world for a while: that is both even more than the much ridiculed FX-9590 , AMD’s ultimate attempt to make some more of its failed Bulldozer architecture.
Later in this article, you will see that the effect differs slightly between AMD and Intel, but in general there are two basic principles to make the processor more efficient. The first of these is tweaking the power limits, or the amount of power that a CPU can use continuously. The other option is undervolting. Whereas with traditional overclocking the voltage was increased to enable higher clock frequencies, with undervolting it is actually lowered to make the CPU more efficient.
Setting a power limit limits how high the processor can clock; the automatic boost algorithm is more likely to hit a wall. Undervolting, on the other hand, does not necessarily lead to lower clock speeds. In fact, because more room is created in the power budget in combination with a power limit, an undervolt can actually lead to higher clocks in practice. However, if the voltage is too low, the processor can become unstable, so the stretch in this is not great and can vary depending on how ‘well done’ your individual processor is.
The theory: power limits and undervolts
A year and a half ago I already wrote an extensive article about power limits, then based on a Core i9 11900K CPU. For those who no longer have that matter in mind, I will summarize it again (somewhat) briefly.
The starting point for the power limits is the thermal design power , or tdp. That used to be an excellent indication of the maximum power consumption of a processor, as it actually still is with video cards, but in 2023 Intel and AMD are a lot looser with that concept.
Power limits at Intel
Intel calculates the tdp based on the power consumption at the base clock rate. In the case of the Core i9 13900K, that is 3GHz. In practice, the processor practically never runs at that clock frequency; idle it clocks back further, under load it turbocharges a lot further.
In technical documentation and the bios, Intel also refers to the tdp as PL1. In addition, there is PL2, a higher power limit that the processor is allowed to use for 56 seconds while boosting. However, since Alder Lake, twelfth generation Core, the overclockable K models are allowed unlimited use of PL2, effectively making PL2 the only relevant power limit for these processors.
The PL1 and PL2 power limits are completely freely adjustable, also on, for example, a B-series motherboard and a non-K processor. So de facto you can make the duration of the maximum boost infinite on every model and that is even the default setting on many high-end motherboards. Naturally, the processor will still throttle if the temperature rises too high.
Power limits at AMD
AMD admits frankly that the tdp does not really have a hard definition; the manufacturer sees the tdp as a ‘class designation’ and not as the maximum power consumption. If the cooling and motherboard VRM allow, a Ryzen processor can consume more than its tdp.
35 percent more to be exact; that’s the difference between the tdp of a Ryzen cpu and the package power tracking , ppt for short. In practice, the ppt is the real power limit of AMD processors, although it is not fully used by every model. Take the Ryzen 9 7900X and 7950X, for example, which both have a tdp of 170W and therefore a ppt of 170*1.35=230W. In our tests , the 7950X eats up almost all of that, but the 7900X, which counts four cores less, needs 187W.
In the bios of an AMD motherboard you will find two other power limits in addition to the ppt: the thermal design current , for long-term load, and the electrical design current, for peak load. These are not specified in watts, but in amps. In practice, you usually run into the ppt limit first.
A term that you sometimes encounter with AMD processors in this context is ‘eco mode’. This checkmark applies the power limits that apply by default to a 65W processor, i.e. with a ppt of 88W, to each processor. Its operation is no different than manually setting a lower power limit as we will do on the next page, but by choosing a value yourself you naturally have more control over the end result.
Undervolting is simply lowering the voltage at which your processor operates. By default, that voltage (the Vcore) is variable, because at low clock speeds/idle, a processor needs a much lower voltage than during a heavy calculation task where all cores work at maximum speed. Ironically enough, the Vcore standard also tends to collapse. To compensate for that, all modern motherboards feature load line calibration .
You can basically set a fixed Vcore for all clock speeds, but it’s a better idea to apply an offset from the standard v/f curve, the curved line that couples an applied voltage to every possible clock speed.
With an Intel processor you do this by choosing an ‘adaptive Vcore’. You can then set an offset for various points on the v/f curve, for example -0.030V for multiplier 58. In the case of the 13900K, that multiplier 58 corresponds to the maximum single core boost of 5.8GHz.
At AMD you can choose Precision Boost Overdrive 2 since the Ryzen 5000 processors. Like Intel’s adaptive Vcore, this feature does not manipulate the v/f curve equally at every point. To do this, enter a number of ‘counts’ in the bios. A count represents an undervolt of 0.003V at heavy load and 0.005V at light load. Most of the users in the Zen 3 overclocking thread report a maximum undervolt between 15 and 30 counts before their processor becomes unstable.
Get started with power limits
Time for practice. The systems I worked with looked like this:
AMD Ryzen 9 7950X
Intel Core i9 13900K
ASUS ROG Crosshair X670E Hero
Gigabyte Z690 Aorus Pro
Corsair Vengeance RGB 32GB DDR5-6000
Nvidia GeForce RTX 4090 FE (only during game test)
NZXT Kraken X62 @ 12V
Samsung 970 EVO 1TB
I first tested both processors at stock settings and then set them to various power limits. You can set the power limits accurately to the watt. So I have selected the power limits chosen here purely to be able to compare the CPUs fairly. You don’t necessarily have to go for one of these values yourself, but you can also choose an intermediate value.
AMD’s 7950X typically consumes 235W; the 13900K even goes to 334W without limitations. It is striking that the Intel processor adheres exactly to the set limit, while the AMD chip tends to exceed it by a few watts.
The fact that the processors are actually clocked much too high by default is evident from the performance decrease, which remains very limited during the first reductions in power consumption. The 7950X still achieves almost 98 percent of its original performance level at 190W, and even at 150W, it’s still close to 93 percent. That while the processor only consumes 66 percent of stock.
The decrease from 334 to 253W, its official tdp, means a 24 percent reduction in power consumption for the 13900K, while the performance drop is less than 5 percent. However, with the same power consumption, the AMD chip often performs better and is therefore more efficient. Only at a power limit of 80W do they end up almost equal.
When power consumption decreases, heat production also decreases rapidly. At 190W, the AMD and Intel CPUs become 76 and 63 degrees respectively. Those are certainly comfortable temperatures compared to the 90+ degrees we note on stock settings.
Now Cinebench R23 Multithreaded is a workload that can max out all cores, but of course a lot of other software doesn’t do that. For example, many games still use a limited number of cores and do not overload them to the max. What do you notice about those power limits?
The graphs below based on our tests in the game F1 22 make that immediately clear: nothing at all. We do not measure any difference between the stock settings and even the lowest power limits of 80W, both in frame rates and frame times. This is because the processors under this load already consume only a fraction of what they need in an all-core load such as Cinebench. Many of the power limits tested above therefore have no effect during gaming and although the AMD chip in particular with an 80W power limit becomes a bit more economical and cooler, this has no measurable effect on the in-game performance.
Now, it might be a bit shortsighted to put this performance image on any workload that isn’t rendering, but it’s exemplary of what you see in a lot of everyday software. The exorbitant power consumption and associated heat production of the top CPU models is only a problem for tasks that can also push such a processor to its limits. With lighter use, the processor does not run into set power limits, or to a much lesser extent, and the effect of this is therefore much smaller.
Get started with undervolting
As mentioned, undervolting AMD and Intel processors works slightly differently. With the Ryzen 9 7950X I use the Curve Optimizer option that is part of Precision Boost Overdrive 2. The maximum stable undervolt was -25, which corresponds to -0.075V under heavy load and -0.125V under light load. However, this may vary by processor. If you really want to push the limits of your CPU’s capabilities, it’s even possible to specify a different number of ‘counts’ per core.
With Intel, you set an offset directly for different points on the v/f curve, which always correspond to a certain ratio (multiplier). For example, V/f-point 9 corresponds to multiplier 58 and is therefore the voltage that a core receives when it runs at 5.8GHz. A fixed undervolt has a proportionally greater impact at a lower ratio, so to prevent any instability, I have only adjusted the v/f points 5 and above for now.
The maximum stable undervolt at the 13900K was -0.030V. That may seem little compared to what was possible with the AMD chip, but in practice more seems to be happening. The voltage during a Cinebench run drops by 0.132V to 1.188V. Now undervolting there are always more roads leading to Rome, such as using LLC to deliberately introduce Vdroop, but to illustrate what is possible this will do just fine.
Now something you didn’t expect might happen; the processors don’t get slower from an undervolt, but faster. That is as follows. With a lower voltage, the power consumption drops, which the processors can use to their advantage. In the case of the AMD Ryzen 9 7950X, the CPU uses the freed up space in the 235W power limit to boost higher. You see the clock speed increase by more than 150MHz, while power consumption and temperature remain practically the same. This leads to a performance improvement of more than 3 percent.
The Core i9 13900K already clocked maximum, so we see a huge effect in power consumption and therefore heat production. The CPU is 85W more efficient and 18 degrees cooler. The difference in performance is almost negligible at 0.9 percent.
Again, the performance increase with the AMD chip is purely because the standard power limit is saturated and the undervolt thus frees up budget within that power limit. This will not be the case with all tasks that do not put such a heavy load on the processor – such as web surfing, photo editing and gaming – and the undervolt will therefore only have an effect on power consumption and temperatures.
Combine power limits and undervolts
The best is yet to come, because you can also combine both techniques to make the processor more economical. For this I have always set the settings used on the previous two pages, an undervolt of -25 and 0.03V respectively and power limits of 80 to 253W, at the same time. In the graphs you will also find the results of the test on standard settings, only the undervolt and only the power limits, so be prepared for a lot of bars.
It is immediately noticeable that with the Intel processor almost the highest performance results from this combination. We record 40,739 points with a combination of a 253W power limit and the undervolt. That is faster than stock and practically as fast as stock with undervolt. At AMD, the latter configuration narrowly wins, although 190W with undervolt is only 1.6 percent slower.
Even at lower power limits, we see the processors with undervolt perform significantly better. For example, at 80W, the 7950X and the 13900K both get 6 percent faster with undervolt. The reason behind this is basically the same as why the 7950X on the previous page got faster from undervolting. You can see that most clearly here in the voltage and clock speed graphs. The undervolt leads to a lower VCore voltage, which basically reduces power consumption. This frees up space within the set power limit. It is used by the boost algorithm to increase the clock speed and that naturally makes the cpu faster.
Because the space in the power budget is used directly to clock higher, you see little or no effect on the actual power consumption and the measured temperatures. They depend almost entirely on the set power limit.
The ultimate goal: efficiency
The ultimate goal of tuning your processor to be more economical is efficiency, or the best possible ratio between performance and power consumption. Of course you don’t buy a high-end processor and then tune it in such a way that you get the performance of a Core i5 or Ryzen 5, so where the sweet spot lies for you, you will have to decide for yourself. In the table below, I calculated the efficiency of each tested setting by dividing the achieved Cinebench score by the actual power consumed to get a score in Cinebench points per watt.
Both AMD and Intel processors are least efficient on their default setting. The 13900K is the furthest of these two, with an efficiency of only 121 points per watt. With an undervolt, the efficiency immediately increases considerably, by 36 percent to 164cb / W. That’s similar to what a 7950X does on stock. With lower power limits, the 13900K slowly creeps towards the 7950X, to eventually work more efficiently than the AMD CPU in the 80W tests. Already in the 125W tests, the 13900K is more than twice as efficient as on stock.
Even up to the 80W tests, the efficiency of both processors continues to increase, indicating that the processors are clocked well above the most efficient point on their v/f curve by default. Although the CPUs work a lot more efficiently, they are considerably slower than standard. In the table below I have compared the performance, power consumption and efficiency with stock, so that you can determine for yourself where the sweet spot would lie for you.
I personally think that 90 percent of the original performance is a good goal. With the 7950X you only achieve that with a 125W power limit and undervolt. The processor then consumes less than half compared to stock and is 68 percent more efficient. The 13900K gets to that point a lot earlier; at 190W and undervolt it reaches 90 percent of the original score for the last time. With 57 percent of the original consumption and a 58 percent higher efficiency, it is still clearly less economical than the 7950X. If you want to end up with less than half of the original consumption, just like with the AMD CPU, then you have to go for the 150W option, but then the performance loss increases to 15 percent.
Now that I look back at the graphs and tables in this workshop, I have mixed feelings. On the one hand, it’s impressive to see how much more efficient you can make a modern high-end processor. On the other hand, most users, even most users, will just run the CPU as it came out of the box. And then the power consumption under a full, all-core load is significantly higher for what is ultimately a fairly marginal performance gain.
The test results have proven that your processor does not have to slow down at all due to the introduction of a power limit and an undervolt. To highlight the highlights; the AMD Ryzen 9 7950X can still deliver 91 percent of its original performance at 54 percent of its original power consumption. And with only a quarter (!) of its original power, the Intel Core i9 13900K still scores almost 70 percent of its stock result.
The fact that these processors are tuned so inefficiently by default has everything to do with how competitive AMD and Intel are at the moment. For much of the last decade, Intel in the top segment had nothing to fear from AMD and therefore no reason to ramp up its products to the maximum. AMD’s competition has led to this in recent years and you can hardly argue other than that AMD has chosen the same path with its Ryzen 7000 processors. Everything to have the longest bars in the benchmark charts.
And yes, that competition has brought us a lot; in seven years, the fastest ‘mainstream’ desktop CPU has become seven times faster. That would never have happened without the resurrection of AMD. The fact that AMD and Intel are so evenly matched led to a significant price reduction of the brand new Ryzen 7000 CPUs after just two months . As a consumer, that makes you happy anyway, of course.
Still, although users once started as a site with overclocking as the number one national sport, it may now be time to say to each other: it could be a little less. That extra hundred watts for the bottom of the performance jug, let it sit for once. And then you may no longer look so fondly at the Cinebench score that appears on your screen, but instead you look at your energy meter with a grin.