Why intel processors draw more power than expected: TDP and Turbo explained
anandtech.com
anandtech.com
I think the basic problem is that power management on modern CPUs has gotten complex enough that you can't really summarize a CPU's power consumption in one or two numbers anymore. To really get a clear picture takes a full blown datasheet with tables showing power consumption under a range of conditions. Frankly, most enthusiasts aren't equipped evaluate such detailed information. Instead people rely on a few sparse data points provided by tech review sites and generally just throw a really big cooler on their CPU and hope for the best.
Enthusiast users building desktop systems tend to massively over-provision both power supply and cooling anyways so it's usually not an issue.
And if you don't want to spend money, you don't want to over-provision either.
...and so does every desktop user, no? If the cooling system can "take the heat", then the CPU should run as fast as it can. It seems like the whole point of this premature throttling is only to meet some stupid marketing spec, and the motherboard manufacturers are wise enough to ignore it.
At my last company we spent extra to buy desktop cases with extra but quiet cooling, so we had speed and quiet at the same time.
I think both Intel and the motherboard manufacturers are to blame: Intel for using a marketing number that won't be hit under most configurations, and the motherboard manufacturers for having over-agressive and unexpected defaults.
The issue is what happens to benchmarks, because removing the power limit makes the limit thermal, which makes performance proportional to the efficacy of the cooling solution and other variables like the ambient temperature in the room.
There will also be variation between individual processors of the same model, since it's always been the case that some would run warmer or cooler, but now that affects out-of-the-box performance. (And with manufacturers sending chips to reviewers for testing, which ones are they likely to choose?)
This is a pretty big deal if you're a business wanting to buy the kind of small form factor desktops with cooling solutions that hew close to the official TDP numbers but you're looking at benchmarks done in larger cases with gamer-typical cooling solutions. Or people buying always-on home devices with more stringent cooling or efficiency requirements. Really anyone who wants to know how the processor would perform in an environment that it isn't permitted to continuously dissipate 180W.
There surely is an upper limit? Would cooling the CPU with liquid nitrogen, for example, make it perform appreciably faster even at the same stock clock?
The concern is that a lot of the cooling solutions that are common in the market could be substantially worse than that, or even motherboards that can't supply that much power and are correspondingly configured not to.
It implies that a small form factor machine may be a lot slower even if it has the same processor in it.
(I'm curious if these new CPUs based on smaller processes do have a measurably shorter lifetime in practice; a lot of people can probably tell stories of decade-old Prescotts and such which are still in active use.)
Mystery might be also in the nuance of turbo vs base frequency draw. So even if you design for TDP as the "max thermal dissipation" you might get only base frequency performance.
On a side note people often confuse TDP with the max power draw of the processor. Intel does not specify total power draw probably because processor scaling is such an important consideration. If you have a system, the easiest way to figure out what max draw from the power source is to consult the power brick.
That is really what is happening. If you have '4 core 3Ghz Base 4.5Ghz turbo 95W TDP' what that means is that the CPU will pull, by spec, a maximum of 95W with all cores running at the base frequency under full load. Turbo can exceed this power spec and maximum power draw under turbo is ill-defined.
> On a side note people often confuse TDP with the max power draw of the processor.
It's the same thing, any 'power' a CPU 'draws' is converted to heat, by definition.
> Tau is a timing variable. It dictates how long a processor should stay in PL2 mode before hitting a PL1 mode.
with the Intel one:
> Turbo Time Parameter (Tau): An averaging constant used for PL1 exponentional weighted moving average (EWMA) power calculation.
I'm assuming that this power calculation is calculated this way, in a similar way to how Unix load averages are. But the thing is that they are an exponentially weighted moving average. I understand why: moving averages need the storage of all the values over that time period and the EWMA only needs the last value of the moving average.
It's not exactly clear how tau relates to the constant used for the exponentially weighted moving average. Assuming I want to minimize the difference between the EWMA and the moving average over a time interval, there appear to be many valid ways to do so -- and this appears to depend on the statistical properties of the variable being averaged. But the Anandtech article shows it as if it's a constant time, which is subtly different.
Is this defined in more detail anywhere?
You're falling into the "TDP is power input" trap.. TDP and power input are not the same. You can have a chip that takes less power to operate at a TDP equivalent to another chip which takes more power to operate.
Edit: There is no way that 100% of output power in a CPU is thermal, else it would be nothing but a very expensive heater. Also, CPUs do not operate in a closed environment, like those replying to me here seem to think.
What the article is pointing out that TDP is thermal design power. Meaning a nameplate value that was targeted for "standard" operation, even though the processor can do higher. So if something has a TDP of 100W, yet the heatsink can maintain proper operating temperature while dissipating 150W, then the processor might sit there drawing (and dissipating) 150W the whole time.
As an aside, I really wish CPU cooler manufacturer/reviews would start simply stating the thermal resistance - basically the single number that characterizes its effectiveness. I know it would undermine the dog and pony show of building a test system with so-and-so's motherboard etc running benchmarks, but the entire goal of science is to weed out the extraneous narrative.
No shit. The trap is thinking that all output power is thermal (it's not).
It's probably fairly neglible though, probables less than a watt in most systems.
There's also this: (https://en.wikipedia.org/wiki/Landauer%27s_principle) if you want some light reading.
The "scandal" was caused by the fact reviewers were getting different results and people would accuse people on not doing the benchmark right.
I think from power/watt number on a cpu people expect two thing. How efficient it is for the performance (electric bill) where you need input power. How better a cooling I need will need the thermal power.
But even with that the issue of Turbo makes it harder to quantify those two because it all depends on workload. If I recall correctly though even with all that if I understand correctly the peak power draw in intel gets throttled much easier than AMD whereas their TDP number (compared to AMD) artificially inflates their efficiency reputation. I do think they had been much more efficient than AMD specially pre Zen.
Again, TDP means the same thing now that it did 15 years ago.
It would be nice to know what the expected power input / heat output is in that state, so you can get cooling to match, and have it in turbo mode at all times when the CPU is being stressed.
Cooling to the TDP only will (or should) get you the base clocks and a brief amount of turbo, but is likely to thermal throttle quickly.
This also brings more into question the adequeacy or not of the thermal interface material used between the CPU die and the heatspreader -- if the CPU will run turbo for longer if you keep it cooler, there's more frustration about less heat efficient solutions.