Official Intel 10th Gen Comet Lake-S CPU Power Limits Revealed
Intel’s Comet Lake-S Parts Can Consume Up To 3.5X Rated TDP
With any processor, and Intel CPUs in particular, power limits are a critical aspect to consider when evaluating a platform’s potential performance. These “PL1”, “PL2”, and “Tau” numbers have been the subject of in-depth reporting from respected outlets such as AnandTech and Gamers Nexus, but it can be a very technical and potentially confusing subject.
Essentially, what this power limit thing boils down to is that motherboard vendors do not necessarily follow Intel’s actual CPU power limits, and Intel hasn’t exactly been transparent about those limits. We are still waiting for a full analysis of Comet Lake-S, which was initially released (like the original Skylake) without any architectural documentation. But we now have more information.
Today, ComputerBase has published a story (German language) that addresses these power limits for the new 10th Gen Comet Lake-S processors (with the following quotes translated from German via Google Translate).
“Incorrect support documents and a lack of information about the start left question marks in Comet Lake-S’s energy management . This has Intel finally cleared and supplies for all processors, the exact parameters of the TDP, PL1, PL2 and Tau. There are some surprises.”
Perhaps the most interesting aspect of this report is that “3.5 times the TDP is officially possible”, though this does not mean the 125 watt parts are going to pull 375 watts – but a low-power “35W” i9-10900T part does actually have a PL2 of 123W (3.5x).
“For example, a CPU can no longer only put the factor 1.25 above the TDP for a short time, in the best case scenario this is the factor 3.5, as the example of the Intel Core i9-10900T shows, with 35 TDP has a PL2 value of 123 watts. Together with a suitable mainboard, which possible tau limits can be undermined by Intel’s tolerance, a completely different CPU is created.
But the regular 65-watt models are also allowed to consume a lot of energy, especially if they have eight or ten cores. 224 watts over the officially planned 28 seconds also correspond to a factor of 3.45. Ultimately, the Core i7-10700K is not as extravagant as originally thought, because it may only use five watts more, but over a longer period. However, as already mentioned, this is already levered out on many boards.”
Table image via ComputerBase
Having these numbers will make it easier to limit the CPU power on motherboards that do not handle this strictly according to Intel specification for each processor. For example, many boards – which already have some form of multi-core enhancement enabled by default – have an “auto” power limit that doesn’t really “limit” anything, as these can provide an effectively infinite period for high power (and high Turbo clocks) – making CPU thermals your only real limit.
Some of the seemingly incongruous power numbers – particularly when reviewing products without strictly enforcing Intel’s CPU power limits to the letter (I have absolutely been guilty of this) – now make a lot more sense. Allowing a high PL2 for the entire duration of a benchmark does produce better results, but also creates a significantly higher thermal load and – of course – draws significantly more power.
There is something to be said for reviewing products in an “out-of-the-box” state, and I doubt very many Intel motherboard buyers have researched PL1, PL2, and Tau numbers in order to ensure that such limits were manually enforced. It’s not advertised, it isn’t “correct” by default in most cases, and that’s just the norm.
Bottom line, is it the practice of running a CPU above the official specs for power limits and duration right? I am not going to get sanctimonious here, as I can see both sides of this. On the one hand it certainly puts more of a strain on both a board’s power delivery and on your CPU cooler, but the result is higher performance. It’s basically zero-effort overclocking. And if we insist on enforcing limits on Intel CPUs the same care must be taken with AMD parts – but they don’t work exactly the same way with regard to power limit/duration, so it’s murky at best.
Once again we see why old hardware is always better (especially those boards covered in little jumpers for CPU voltage and frequency). I rest my case.
Sebastian,
One could read this article as suggesting that runtime of benchmark tests should be longer than they might currently be, so that they are not unduly influenced by the ability of a CPU to support higher power levels for longer periods of time.
Tests which have runtimes shorter than Tau would then appear to be much faster than they really are — tests which run longer (to better reflect real life needs) would provide better comparisons, no?
This is where we get into the area where benchmarks don’t tell us more than how a given part does in a given benchmark. Yes, tests with a runtime under 56 seconds – such as an all-core Cinebench R20 run on a high core/thread count CPU – will be taking advantage of the higher power available within even a strictly enforced limit. So if we use Blender instead, that limit will expire long before the render is finished, making that a very good way to see frequency and power over time. Blender will be used here at the site to measure power and frequency going forward.
I just struggle to say that processors are “faster than they really are” when they are designed to boost to higher clocks for a shorter duration. Everyday computing involves a lot of ‘bursty’ workloads. Of course Intel would like it if we just talked about daily usage models, and used benchmarks that showcase short-duration and typical single-threaded CPU workloads… There has to be a balance. It’s awkward that motherboard vendors don’t respect these limits, causing most systems to run in an effectively overclocked state.
Hi everyone
I like the option of stretching the PL2 and tau on non-K CPUs, this is the only way to OC them (using term loosely)
Can you comment on how many motherboards do implement Intel’s guidelines as opposed to not?
Of course I am going to research what the particular motherboard I’m buying is doing (though you’d be nuts to buy Intel now), so I’d almost think that given infinite time and resources you should test in both modes – power capped and uncapped.