Coffee Lake & Older CPU PL1, PL2, Tau Example Specs

For example, the 10900K’s max limited-core turbo multiplier is 53x and all-cores is 49x (when the CPU is operating below 70C). Breaking these turbo limits is considered overclocking. Intel also has guidelines for short-term power draw, long-term power draw, and boost duration. These variables are not part of the enforced spec, so breaking the suggested limits is not considered overclocking, and motherboard manufacturers go hog-wild pushing these limits to the highest values that the assigned bytes can hold, resulting in thermals and performance that are completely out-of-whack with what the processor’s listed capabilities would suggest. For example, the 10900K will run at 4.9GHz under an all-core load forever in many boards, whereas our stock testing with Intel stock limits applied showed the CPU running at 4.1-4.2GHz across all cores after 56 seconds elapsed. 

Recently, we’ve had several of you tweet at us or post comments about how MCE and similar features caused system instability, blue screens, or other troubleshooting challenges, and that’s because it’s effectively pre-overclocked by the manufacturer, just without asking you. Pay close attention to our use of the word “effectively,” here, because it isn’t “officially” overclocking -- but that’s ultimately what’s going on. The clock is higher than expected -- you might even say it’s “over” what the expected value -- and so we ignore semantics of “spec,” “guidance,” and “overclocking,” and instead go with what makes the most actual sense for the user, bullshit marketing labels aside.

05:28 | Some Board Makers Violating Official Guidance

Here’s a preview of the problem--we’ll get into more test results and methodology in a moment, but this sets the stage. Each of these numbers was taken directly from the EPS12V cables, before VRM efficiency losses but also after the wall. The Blender measurement within the 56-second window and at five minutes help show what settings are used or ignored. The difference between using Intel guidelines and using some form of MCE can be the difference between 118W and 244W, as seen briefly here. That’s also the difference in cooler choice, because now you’re dealing with substantially more power, and is further the difference in stability for some systems.

We picked a selection of motherboards from different manufacturers and tested them at default settings using an i9-10900K and the RAM and GPU from our CPU review bench. By “default,” we mean resetting the board entirely to defaults with no changes other than maxing out all fan speeds. This is in contrast to the stock CPU tests we do for reviews, where we set RAM frequency and timings, explicitly disable MCE, and set the correct DRAM voltage, among other things. The latest publicly available BIOS was used for all motherboards except the MSI Ace, where we used a reviewer BIOS to avoid some potential issues outside the scope of this piece, and the ASUS Extreme, where we used the same reviewer BIOS that we did for the Comet Lake CPU reviews.

We used a custom tool to identify how the boards are configured once in Windows, which allows us to ensure all settings are as expected from BIOS.

ASUS’ Z490 boards are the only ones we’ve tested that force the users to explicitly choose between MCE or Intel defaults, and default to the Intel recommendations when the board is reset, at least in the BIOS revision that we used for our Comet Lake reviews. This screenshot [above] was taken on the Extreme, but the readout for the Hero looks exactly the same. These are the settings we want for benchmarking within Intel’s official guidance. Going outside of this is not standard and would not be an accurate representation of Intel; exiting official Intel guidance, if it’s done, means you might as well also throw PBO onto the AMD parts. It might be semantically called “spec” to run outside of guidance, but it is not within reason and it is a practice which could yield results which do not accurately represent all parts tested.

This screenshot was taken on the MSI MEG Z490 Ace, but the Gigabyte Z490 Aorus Master had similar settings, with the turbo ratios and per-core max ratios unaltered and PL1, PL2, and Current settings maxed out. Note that the reduced power limit duration is effectively meaningless since PL1 and PL2 are both set to 4095W, which we’re never going to hit -- even with liquid helium. Based on these settings, we can predict limited-core boosting up to 5.3GHz as usual and 4.9GHz sustained on all cores under load for as long as the temperature allows.

This final screenshot is from the ASRock Z490 Taichi, which had the wackiest default behavior of all. The multipliers for active cores are still unchanged, but the multiplier per core is unlimited, meaning that this board will behave similarly to the Ace and the Master, but with any two cores allowed to boost to 5.3GHz rather than only the best cores, which are cores 8 and 9 here. This isn’t an official tool, so we’ll check these numbers with HWiNFO in a moment to make sure it’s not a bug.

07:17 | Thermal Impact of Board Choices

Now that we know what settings are applied by default on each board, we need to take a look at just one of the end results of those settings. We use a Kraken X62 CLC at max pump and fan speeds for all desktop-class CPU testing, and you can learn more about our CPU test bench component selection and methodology in our standalone test methods piece. The temperatures getting charted now were calculated by averaging the per-core temperatures and taking a delta against the ambient room temperature, which was measured every second of the test. The truncated scale is to help illustrate the change on a line plot, and tops-out at 50 because it’s dT over ambient. The 10900K installed in the Gigabyte Master, which we already know from our first chart consumed the most power during this test, reached steady state at about 49 degrees Celsius over ambient, while the same CPU in the more conservative ASUS Hero levelled-out around 22 degrees Celsius dT. Note that there are significant differences between all of the boards: even the two that obey stock limits are significantly different. Anyone running the 10900K at default settings in the Taichi, Ace, or Master will conclude that it’s a much hotter CPU at stock settings than we know it to be from our review. It’s hard to overstate how huge a gap 27 degrees Celsius is at steady-state -- that’s orders of magnitude higher than what we see in cooler reviews. 

If we put a CPU cooler chart on the screen from our Liquid Freezer II review, it becomes clear that the cooler itself matters a lot less than the power going into the CPU; one, after all, is what sets the requirement for the other.

Back to our thermal line plot, it’s clear where Intel gets its “hot and power-hungry” meme status online, and that’s from its motherboard partners. Although the other boards run at a higher frequency, they also run at over-aggressive voltage settings.

09:52 | VCore Impact

For a clue as to what’s going on, we can next look at vcore as reported by HWiNFO; to really do this properly, we’d need a DMM hooked into each board, so this is a quick spot-check using software. Unfortunately, ASRock’s board reported an obviously incorrect number, and HWiNFO hadn’t yet labelled a vcore entry for the MSI Ace at the time of testing. We could go back and log with a DMM or scope, as we have for past pieces about MCE, but this is enough to get the point across: the ASUS Extreme reports a vcore of 1.128V under all-core load, then drops sharply to 1.012-1.03V after the 56 second turbo limit expires. Even with whatever error exists in the software approach, the point is clear: There’s a firm drop as the need to maintain a stable, high frequency is also reduced. The Hero follows the same trend, but with a slightly lower auto vcore after the time limit. The Gigabyte Master, on the other hand, reported exactly 1.272V throughout the entire test. That’s 1.272V for an all-core frequency of just 4.9GHz. For perspective, we achieved 5.2GHz on this same CPU with a reported vcore of 1.296V in the Maximus XII Extreme. One of the greatest weaknesses of pushing ambitious default settings is that they must cater to the lowest common denominator and lowest quality silicon--if Gigabyte decides that a 10900K exists that needs 1.272V to run at 4.9GHz all-core permanently, they have to make that the all-core voltage for every 10900K ever to socket into these boards, which is bad news for you. Again, these are software logs of data reported by the motherboards and the hardware current sensor chart at the beginning of this piece is a more objectively accurate depiction of CPU power in watts, but clearly the auto vcore assigned to the 10900K on the Gigabyte board is both higher and sustained longer. This is illustrated clearly in our power logs from the hardware logging, and it’s not the first time Gigabyte has been guilty of excessive Vcore. A piece we wrote in 2017 demonstrated this exact issue. That’s how long we’ve been fighting this fight.

12:25 | Package Power Consumption

Here’s a chart for package power, a data point that’s tracked on all five motherboards by HWiNFO. Again, the caveat about this being a software report by each individual motherboard rather than an external measurement applies, especially since the external measurements reported the Taichi drawing slightly more power on the EPS 12V cables than the Ace; further, note that the pre-EPS12V draw doesn’t account for VRM efficiency losses, so although the EPS12V numbers are considered “good enough” for CPU power measurement, you’d need finer controls for board comparisons. Nonetheless, this method of measurement does appear to correspond directly with the PL1 and PL2 we saw above. Neither of the two ASUS boards hit the 250W PL2 limit in the initial turbo boost, as they didn’t require the headroom and aren’t as abusive on voltage. As soon as the time limit was up both were locked down to the 125W PL1 limit, at whatever settings they required to hit that threshold. The other boards without any such limits reported much higher power levels, with the Master reaching ~237W package power, reasonably close to the 244W check with hardware logging.

Now let’s look at the actual CPU frequencies. This is what dictates performance results in games or other applications, and is the most important takeaway from the point of fair performance comparisons between brands. 

In this test, the 10900K behaved similarly with the Ace and the Taichi, which we’ll plot first. The Taichi was almost constantly at 5300MHz and boosting more than the expected cores. This behavior would boost the result higher than other boards, which makes ASRock look better on charts and is why they’re doing it; at the same time, it threatens system stability in some use cases, as our viewers have voiced, and it also requires a more serious cooling setup and power source. The Ace is similar in its behavior. The Extreme shows us how performance should look. The Aorus Master exhibited the most “stock” behavior here, staying at 5.1GHz for the most part with occasional bursts to 5.3GHz on cores 8 and 9. The two ASUS boards followed the same pattern, but with marginally higher frequencies due to some BCLK manipulation that we’ll discuss later on. The Taichi and the Ace both kept at least one core boosted to 5.3GHz much more often than the Gigabyte or ASUS boards.

Examining a couple boards individually, we can see that the Taichi boosts to 5.3 on any core it pleases, often on three or four cores at once, which breaks even the rules the board has set for itself, since the 3-core active limit on the Taichi is still reported to be 51x. 

In contrast, the Extreme is only allowed to boost to 53x on cores 8 and 9, it only actually uses core 8, and even then it only boosts sparingly. This clearly massively influences performance in an unrealistically positive direction if anyone decided to benchmark the 10900K on an ASRock Taichi board at stock settings. Again, we view the ASUS stock settings as more “correct,” though it too has some wrinkles.

For the all-core Blender test, let’s take the average core frequency rather than the maximum limited-core boost. The Ace, Master, and Taichi all kept the 10900K locked to a steady 4.9GHz through the entire test, as could be predicted based on the MCE settings. 4.9GHz is the maximum limited time all-core turbo, and jacking up the limits will keep that turbo sustained forever and outside of guidance; again, if you’re going to test like this, then it’s unfair to also leave AMD’s power limits in place. Both must be treated the same way, regardless of semantics. On the ASUS boards, frequency quickly drops to just over 4800MHz with a 48x multiplier. We might expect to see 49x here as well, but the CPU is hitting some other limit--during this period the performance limited flag in HWiNFO for “Electrical Design Point/Other (ICCmax,PL4,SVID,DDR RAPL)” is activated. After this point, the two ASUS boards behave differently, with the Extreme sustaining an average clock speed of about 4.26GHz and the Hero sustaining about 4.02GHz. After the first 56 seconds, the limiting factor becomes PL1, so the frequency isn’t directly limited and differences can emerge.

We’ve reported on motherboard manufacturers cheating performance by bumping BCLK in the past. This isn’t what we’re focusing on today, but we point it out here to show that even when manufacturers obey turbo guidelines, there are other ways to fudge numbers ever-so-slightly. The MSI, ASRock, and Gigabyte boards reported a proper, steady BCLK of 100.00MHz throughout the Cinebench R20 single threaded benchmark, with the Master in particular showing zero deviation whatsoever, while both of the ASUS boards fluctuated between 100.4MHz and 100.5MHz in the same test. That means that the 53x max turbo multiplier of the 10900K translates to a frequency of 5321MHz-5327MHz for 56 seconds, rather than the 5300MHz it should be; however, please note that we manually configure BCLK to 100.00 for all CPU reviews, and so this has no impact on our results for the review. 

19:23 | CONCLUSION

As reviewers, we find MCE unacceptable. This is up to interpretation for every reviewer, and there’s no objective “best” way to do it. We can see the argument that, in the least, “this board ships with these settings, and so our test is representative of this board and that CPU.” That’s valid, assuredly, and we’re fine with it. The most important thing is that viewers understand this, though, and that they’re made very aware of the potential downsides of using the board out-of-box that way if violating power and turbo limits. At GN, the two reviewers here (Patrick and Steve) don’t personally believe it’s the “right” way to represent the product, because our belief is that it should be tested against its marketing claims, not against changes the board makers made. That’s more of a board review, in our opinion; that said, again, we can understand why different reviewers do it different ways, and we respect those decisions. Our biggest hang-up is making sure that viewers understand the discrepancy, and after that, it’s a desire that Intel puts its foot down for board manufacturers to adhere to “official guidance,” which is functionally “spec,” semantically correct or not.

At best, allowing the board to override guidance can make it difficult to define “stock” performance of a processor, and at worst, it leads to publishing misleading results, particularly when another vendor with its own rules is in play. We already have OC benchmarks, so we don’t need “sort of OC” benchmarks from the board. Now you’re playing with fire as to whose guidance can be violated and when, or whose CPU is allowed to run outside of “technically stock,” while the other one has to stick to “strictly stock because they force motherboard companies to the same settings.” This is doubly true for this generation of processors, where the 53x/49x TVB multipliers are so far removed from the 42x/41x multiplier that the processor actually runs at under sustained load when obeying Intel guidelines. As an example, if we were to forget to disable MCE or forget to double-check Intel’s recommended limits, we could conclude that the stock 10900K is much hotter, much more power-hungry, and much higher-performance than it really is, which is why we’re careful to adhere to the actually advertised specifications for the review. Intel and AMD both take shots at each other in review guides, as we discussed before, with each suggesting that the other have its limits in place. Intel wants PBO off, AMD wants Intel’s suggested power limits applied. Both are concerned the competitor scores higher, obviously, but our concern is fairness, and to us, that means following the advertised specifications.

The arms race between motherboard manufacturers means that it’s unlikely there will be any self-motivated change from that quarter, so real change requires enforcement by Intel. That is to say: these “guidelines” should be part of the spec, more so than they are now. Advertising a tame TDP, ergo advertising a tame power budget (whether directly relatable or not), while allowing high-end motherboards to essentially auto-OC the CPUs is a sorry attempt at having their cake and eating it too. Setting MCE to “default” or “auto” should mean “off,” not “who knows if it’s off or not.” It’s insane that PC builders have to jump through a hoop to explicitly enable XMP, because oh no, your RAM might be unstable, but you simultaneously don’t have to enable MCE, which definitely causes problems.

Motherboard manufacturers have been cheating, are cheating, will continue to cheat without intervention. No, default BIOS settings that exceed Intel’s guidelines aren’t against the rules, don’t technically break spec, and aren’t considered overclocking, but it misrepresents both the CPUs and the motherboards affected. They’re running “over” the normal “clock,” yet it somehow isn’t “overclocking.” We’re all for the ability to overclock, and in some cases we might even be for MCE, but having it as an option that’s toggled on by default is irresponsible. All boards should run CPUs (and every other component) at safe, stable settings until the user specifies otherwise. Watch our recent video on the i9-10900K’s “high” power consumption for more thoughts on this topic.

Editorial, Testing: Patrick Lathan
Host, Test Lead: Steve Burke
Video: Keegan Gallick