Intel Core i9-12900K Alder Lake Tested at Power Limits between 50 W and 241 W

[BSim500]

Doubt it, but will test to verify

I'm really not in the market for a new CPU, but I just want to say I love the battery of in-depth tests TPU has been doing recently plus the level of communication with your readers that puts other sites to shame.

 

[bobmeix]

Great review, but these graphs seem to contradict the values from the 1080p page of the game tests, where the average values are much less divergent.

 

[bug]

I was going to limit my 12600k (when I'll be able to get the mobo), this shows you can shave off the last 50W with basically no impact whatsoever. I'm going to look into setting PL1 even lower than 190W. Despite the first page saying for AL PL1=PL2 and tau=infinite, there's clearly a difference when running PL1 != PL2.

 

[Solid State Brain]

By the way: on Rocket Lake strange throttling behavior can occur if PL1=PL2, and both are set to a low level which would make frequency decrease below the base clock. At least this is what I observed on my i9-11900 (non-k).

I thought it might be worth demonstrating this behavior during a Blender 3D render; it could potentially be affecting this i9-12900K TDP review to some extent. After revisiting it, perhaps it is not strictly related to the base clock like I initially wrote, but it needs more testing to make sure. It does seem to occur mainly at low power limit settings.

PL1=PL2=65W, Tau=1s, reported package power = 65W. See how unstable core clocks are, varying randomly between base clock (2.5 GHz in my case) and some higher clock:



PL1=65W, PL2=85W, Tau=1s, reported package power = 65W. Core clocks are now stable. PC power consumption at the wall is also a few watts lower (not shown here):

 

[bug]

I thought it might be worth demonstrating this behavior during a Blender 3D render; it could potentially be affecting this i9-12900K TDP review to some extent. After revisiting it, perhaps it is not strictly related to the base clock like I initially wrote, but it needs more testing to make sure. It does seem to occur mainly at low power limit settings.

PL1=PL2=65W, Tau=1s, reported package power = 65W. See how unstable core clocks are, varying randomly between base clock (2.5 GHz in my case) and some higher clock:

View attachment 228213

PL1=65W, PL2=85W, Tau=1s, reported package power = 65W. Core clocks are now stable. PC power consumption at the wall is also a few watts lower (not shown here):

View attachment 228214

Interesting, can you test the first scenario after enabling clamp for PL1, too? I suspect that despite clamp being enabled for PL2, the PL1 setting somehow overrides that.

 

[InVasMani]

Intel made it clear several times that this is the way the product is intended to be used.

Intel has chosen to kind of market it as a performance leading chip at all costs TDP be damned. I do think AusWolf makes a alright point though TPU has covered Alder Lake in quite a bit of depth. It seems like every time I turn around there is another article on it covering it further.

I feel what comes next with big LITTLE will be great based on how Alder Lake has turned out. I believe it will mature like fine wine much the way Ryzen has evolved over time. I don't see CPU innovation and competition slowing in a hurry. It seems like the next 5 years or beyond we'll see plenty of strong CPU competition.

I thought it might be worth demonstrating this behavior during a Blender 3D render; it could potentially be affecting this i9-12900K TDP review to some extent. After revisiting it, perhaps it is not strictly related to the base clock like I initially wrote, but it needs more testing to make sure. It does seem to occur mainly at low power limit settings.

PL1=PL2=65W, Tau=1s, reported package power = 65W. See how unstable core clocks are, varying randomly between base clock (2.5 GHz in my case) and some higher clock:

View attachment 228213

PL1=65W, PL2=85W, Tau=1s, reported package power = 65W. Core clocks are now stable. PC power consumption at the wall is also a few watts lower (not shown here):

View attachment 228214

The CPU utilization on the top looks very reminiscent of the way baud rate. One encouraging thing about that is it looks like speed step is pretty quick, but yeah that utilization is just all over the place. You might try tinkering with speed shift bumping up the minimum in throttlestop and see what happens. That might make it more consistent. Try setting the minimum to 44 with 52 max on speed shift and re-test just to see what happens in terms of utilization.

 

[Solid State Brain]

Interesting, can you test the first scenario after enabling clamp for PL1, too? I suspect that despite clamp being enabled for PL2, the PL1 setting somehow overrides that.

Unexpectedly, similar behavior. Depending on load conditions, clocks may spontaneously stabilize afterwards, but by setting first PL1=PL2=low and only then starting a high-load multithreaded benchmark, clocks can reliably be made unstable for prolonged periods of time.

 

[londiste]

@W1zzard, any plans to do head-to-head comparisons with Zen3? P-cores only at 88W/141W to match 5800X?

 

[Solid State Brain]

The CPU utilization on the top looks very reminiscent of the way baud rate. One encouraging thing about that is it looks like speed step is pretty quick, but yeah that utilization is just all over the place.

I think the behavior with PL1=PL2 might be undefined, and the CPU's power throttling algorithms might have some internal processing delay that can cause core clocks to be unstable at low power limits, when set that way.

To be sure that the low-power results in this review are correct, it would be probably best to check if the same occurs also on Alder Lake, but I think the best way to mitigate the issue is setting PL2 to a slightly higher level than PL1 and setting Tau to a very short time (EDIT: for example 1 second or something less. If it's much shorter than this, similar problems start occurring again).

You might try tinkering with speed shift bumping up the minimum in throttlestop and see what happens. That might make it more consistent. Try setting the minimum to 44 with 52 max on speed shift and re-test just to see what happens in terms of utilization.

No change, it seems.

 

[TheoneandonlyMrK]

Awesome test, thank you!

It shows exactly my point when I straight away dismiss claims saying "Intel is inefficient and runs hot". With custom PL settings, it's as efficient as the user wants it do be. Dropping PL values to reasonable levels gives the user a barely noticeable difference in observable performance, but a huge gain in efficiency and heat output. Reading reviewers that only test at Stock out power limits gives a one-sided and unrealistic picture.

Fixed that for you.

 

[InVasMani]

I think the behavior with PL1=PL2 might be undefined, and the CPU's power throttling algorithms might have some internal processing delay that can cause core clocks to be unstable at low power limits, when set that way.

To be sure that the low-power results in this review are correct, it would be probably best to check if the same occurs also on Alder Lake, but I think the best way to mitigate the issue is setting PL2 to a slightly higher level than PL1 and setting Tau to a very short time (EDIT: for example 1 second or something less. If it's much shorter than this, similar problems start occurring again).



No change, it seems.

View attachment 228229

Looked more stable in the first screenshot. What about if you turn off speed shift? Strange scenario I'll give you that if only it sounded as sexy as it looks like a EDM music LFO automation. Is this a big issue with Rocket Lake or can you just bump up PL2 to like 66W or something and it works fine and just a strange bug you encountered!?

So this PL1/PL2 situation with Alder Lake this raises a good question in regard to Intel do you think they'll add a few more stage of PL settings beyond the two in the near future!? I feel like they could double them with a PL1 and PL2 for both E core and P core designs to control and dial in base and boost TDP's on each core type.

Do these power limits have any impact on the IMC!!?? Like can you limit them and get better IMC performance/stability perhaps by throttling the CPU a bit intentionally!? I can't help wonder can they help any with either pushing Gear 1 performance higher or better memory stability and frequency support perhaps out of populating all four DIMM slots for example. It's sort of a shot in the dark I suppose, but wouldn't hurt to try. Explore all possibilities I believe is the saying.

On another note and one that could have a impact on high TDP CPU's I've been considering and pondering a kind of a innovative take on cooling which I haven't seen tested and explored. It may work and prove useful and be good in practice on some of these newer more densely packed harder to cool CPU's at the same time. I spoke briefly about it once earlier in a thread, but I don't know if anyone actually saw or read about it.

So my thoughts were what if you paired a Noctua NH-D15 heatsink tower cooler and in between where the heatpipes connect to a flat base what if you adhere and mounted a water cooler block to it!!? Then run it to a radiator up to!? Any thoughts on how it might impact temperature do you think it would transfer heat well enough to the water block mounted on it and could then transfer that to the radiator!!? Now say it can you could go further with have it run from one pump to another Noctua NH-D15 mounted somewhere else so two water pump blocks and two Noctua NH-D15 heatpipe tower coolers in place of a radiator essentially.

You could even include radiators though to be fair as well actually if utilizing two or more water block pumps w/o really slowing down the water flow if you manage the line lengths in smart way it might even increase the water flow. It comes down to how far the line lengths are essentially. In the case of skipping the radiators entirely it would be pump to pump and be like two network adapters with teaming in terms of the water flow increasing the pumping power between two Noctua NH-D15 cooler's one directly mounted to the CPU the other mounted wherever you feel like CD-Rom bay perhaps?

So anyway thought I'd share that hypothetical cooling idea. I haven't seen anyone do it or consider such a approach off hand so I'm really curious. What's really interesting about it is it could be nice and silent. It could be dead silent in fact outside of pump noise or put your fans on the heatpipe tower's and basically double down with a bit of water cooling heat exchange transferring between them in sort of a NIC teaming inspired manner of speaking.

 

[Solid State Brain]

@InVasMani : SpeedShift was not enabled when first I observed and reported this strange behavior. It is not a big issue in practice, since in most cases, including low-power devices, PL2 will be typically set to a higher level than PL1 (even if just for a few seconds or less) and if PL2 is larger than PL1 the bug does not show up.

It is only an issue if you want to limit power to a specific level for benchmarking purposes; then, you have to make sure that with PL1=PL2 the CPU is operating correctly, otherwise performance and system power consumption might not be consistent with the reported package power.

 

[Wirko]

It's interesting seeing the efficiency chart show a sweet spot of 75-100w. Surprised to see 50w rated worse.

At 50W, the processors seems really pushed out of its comfort zone - I mean, the range in which it can adapt efficiently to the limits imposed.

Given how hairy the frametime graphs become at 50W, I'd even suspect that some throttling occurs, not for thermal reasons but to save power.

Edit: Seeing @Solid State Brain 's measurements, I'd rather say "even more throttling" than "some throttling".

 

[bug]

@W1zzard, any plans to do head-to-head comparisons with Zen3? P-cores only at 88W/141W to match 5800X?

This is probably the last of our Alder Lake special investigation articles.

 

[Vayra86]

Great articles, I'm still surprised mostly at how 'all over the place' Alder Lake's performance really is.

Chart top and bottom all in one product, in any combination of metrics. It also shows how unnecessary it is, in the same way. You could get all this performance already for quite some time, and where it does offer more, you probably never needed it.

Intel should have marketed this with 125W peak, really, and not PL1=PL2 just to win synthetics.

 

[bug]

Great articles, I'm still surprised mostly at how 'all over the place' Alder Lake's performance really is.

Chart top and bottom all in one product, in any combination of metrics. It also shows how unnecessary it is, in the same way. You could get all this performance already for quite some time, and where it does offer more, you probably never needed it.

Intel should have marketed this with 125W peak, really, and not PL1=PL2 just to win synthetics.

I agree with most of that, but AL is not unnecessary. Sure you could get that performance already, but now you can get it for cheaper (i.e. 12600k is a cheaper 5800X). As lesser models (and chipsets) hit the market, this will become more apparent.

 

[Zubasa]

This reminds me so much of what AMD did with the 5800x - gave it a PPT of 142W, and let it burn at 90C to win cinebench runs.
Cap it to ~120W and suddenly you're 30C lower for 1-2% less performance


125W seems to be the best middle ground to me as well, maybe 150W (which was not tested) would give the best of everything (And in that case, using very similar wattages to the high end ryzens eliminating the complaints about higher cooling needs)

You basically described the 5800 non-X for OEMs. That chip is one of the most efficient Desktop CPUs out there.

@W1zzard, any plans to do head-to-head comparisons with Zen3? P-cores only at 88W/141W to match 5800X?

AMD pretty much did just that.

 

[Raendor]

Awesome test, thank you!

It shows exactly my point when I straight away dismiss claims saying "Intel is inefficient and runs hot". With custom PL settings, it's as efficient as the user wants it do be. Dropping PL values to reasonable levels gives the user a barely noticeable difference in observable performance, but a huge gain in efficiency and heat output. Reading reviewers that only test at maxed out power limits gives a one-sided and unrealistic picture.

My friend, exactly the same case here with majority of noobs or fanboys simply not understanding how PL1 and PL2 work and what those 240w meant for last 3-4 gens or so. They completely don’t realize how much efficiency can be gained from setting your limit based on what you use it and how little k processors offer for home usage.

I personally run my 11700 with 125/140 limits in itx Ncase on air (silver soul 135) with temps not going beyond mid 60s in cpu-tense games and simply laugh at all “inefficiency” talks amd fanboys are so keen on.

 

[Vayra86]

I agree with most of that, but AL is not unnecessary. Sure you could get that performance already, but now you can get it for cheaper (i.e. 12600k is a cheaper 5800X). As lesser models (and chipsets) hit the market, this will become more apparent.

Certainly, but I was mostly referring to how Alder Lake tries to muddy the waters with its PL1=PL2 nonsense, it has absolutely no practical advantage.

Competition wise, this is great.

 

[Xebec]

Excellent article - I've been following TPU for years and years and finally decided to register I really can't say enough about how much I love this website's data and testing methodologies - it's an extremely helpful reference.

One question - will there be a follow-up discussing temps or anything like that for the various wattages run? Temperature can affect the power used independently of the fact that power itself generates heat. I'm just curious academically.

An old 3770K, for example - can increase power usage by 20W just by running hotter - same voltage (1.359V), same clock speed (3.2 GHz), and same application test - the difference being 60C vs 100C operating temperatures. This sort of implies even more scaling losses as you go higher in overall power usage. Interestingly, at a much lower voltage - 0.923V for example, power seemed to only increase ~ 12-15W when the CPU went from ~ 40C to 100C (a wider band too).

Sources:

i7-3770K vs. i7-2600K: Temperature, Voltage, GHz and Power-Consumption Analysis

It has been just a little over a year now since I posted my deep-dive analysis of the dependence of power-consumption with temperature, clockspeed, and cpu voltage for my 32nm i7-2600K cpu (sandy bridge) and I felt like it would be a nice evolution to perform a similar analysis of my 22nm...

forums.anandtech.com

https://i272.photobucket.com/albums/jj163/idontcare_photo_bucket/Intel%20Core%20i7-3770K/i7-3770kat32GHzPowerConsumption.png

 

[Ed_1]

W1zzard: on the Wprime 1024M tests are you running the test with all threads enabled?
Cause my 12600k with a 16 thread count setting scores 79.911 sec

When I start Wprime 2.10 it defaults to 4 thread count.

 

[HD64G]

Any chance of doing the same type of efficiency test with Ryzen 5900X or 5950X? That is even easier. Just needs changing the power limit from auto (142W) to 105W, and another step (75W maybe?) or 2 down to tis lower limit @45W.

 

[bug]

Certainly, but I was mostly referring to how Alder Lake tries to muddy the waters with its PL1=PL2 nonsense, it has absolutely no practical advantage.

Competition wise, this is great.

I concur to a degree, but honestly, mortals don't know what PL1 and PL2 are.
The only practical effect is these will run as hot as the cooler allows (reviews use high-end cooling, end users do not, typically) and thus rather loud. If Intel doesn't dial PL back in locked chips, these will end up next to Netburst in history.

 

[Solid State Brain]

The issue I have with the default PL2 setting used until Rocket Lake, when it's intended to be much higher than PL1, is that the default Tau of 52 seconds (or even 28 seconds on lower SKUs) does not make much sense if the aim is just taking advantage of the thermal inertia of the cooler. After so many seconds at high power, temperature can easily increase close to the thermal limit, and fans will easily respond to that. For this mode of operation to make sense as claimed, PL2 would have to last just a few seconds.

A different way of using different PL1/PL2 advantageously is disregarding thermal inertia and setting PL2 to a high level so that short-term intense multithreaded loads can take place at full speed. Then, PL1 can be set to a low level after a Tau time of several minutes for more efficient operation for long jobs (3D/video rendering, etc).

Default settings resemble more the second scenario, but then a 28s/56s Tau feels short for that.

 

[dicktracy]

Just setting PL2 in the BIOS doesn’t help much since the CPU and motherboard will still use high voltages for no reason. Best method is setting PL2 at 241w and reducing the voltage via V/F and you’ll get same performance at around 170w CPU package in Cinebench. 27k multi points at 170w is efficient as hell.