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

[W1zzard]

I'd be curious to know whether there's a setting for tau as well in the UEFI.

Yes there is, but when running PL1=PL2 it doesn't matter

 

[Fifagi]

In asus z690 how i can set pl1=125 pl2=125?

Thanks n

 

[W1zzard]

In asus z690 how i can set pl1=125 pl2=125?

Thanks n

set long duration and short duration power limit to 125 W

 

[InVasMani]

What are you on about?

https://www.igorslab.de/en/up-to-60...vidias-rtx-3090-ti-makes-the-start-exclusive/

 

[bug]

Yes there is, but when running PL1=PL2 it doesn't matter

Obviously, but as stated above, I'm planning to set PL2 to 190 and try a lower PL1

 

[THANATOS]

Many thanks for the very interesting review.

What I find weird is that i9-12900K limited to 50W for both PL1 and PL2 provides only 8872 points in CB R23 and the power consumption is 120W.
If you check Intel Core i9-12900K E-Cores Only Performance Review then you will find out that i9-12900K with only E-cores enabled provides 10366 points in CB R23 and consumes 118W.
You gain almost 17% higher score and power consumption is actually 2W lower.

The other interesting thing is the actual clockspeed of this i9-12900K limited to 50W.
If i9-12900K's P-cores work at 4.9GHz and E-cores work at 3.7GHz with PL1=PL2=241W and generate 27780 points in CB R23 then PL1=PL2=50W and score of 8872(or ~32% of Full performance at 241W) would mean P-cores are working at measly ~1568Mhz and E-cores at 1184MHz.

 

[AusWolf]

Performance in multithread benchmarks like Cinebench should increase almost perfectly linearly with frequency, but the presence of slower E-cores on Alder Lake makes this relationship less clear.

I tried repeating a test at fixed all-core frequencies from 4.6 GHz down to the base clock of 2.5 GHz with my i9-11900. Plotting the results against CPU frequency makes them easier to understand. I just varied all-core boost frequency, not the PL, so perhaps this has an effect as well, but it is clear that power requirements increase more or less exponentially with frequency, so squeezing a few hundreds more MHz would have a very large effect on them. This should a general principle valid for all CPUs.

EDIT: by the way, the effect of temperature on power consumption can alter the results as pointed out earlier by Xebec, so this would have to be taken into account when doing CPU power testing. Hotter CPUs will throttle at lower frequencies for the same power limit when hot, or conversely consume more power if the frequency is fixed.

Interesting findings. I did a similar run with my 11700, which is (theoretically) identical to your 11900 except for its different boost table and lack of TVB.

At 65 W, your 11900 scored way above 10000 points - my 11700 scored 9267.
At 125 W, yours scored around 14000 - mine scored 11700 (funny enough).
Your diagram ends a bit above 140 W and 14500 points (I guess?) - with unlocked power limits, my CPU eats around 165 W in CB R23 and scores 14100.

With this regard, it would be interesting to see how differently a 12900K and 12900 non-K limited to the same power limit perform.

 

[InVasMani]

Many thanks for the very interesting review.

What I find weird is that i9-12900K limited to 50W for both PL1 and PL2 provides only 8872 points in CB R23 and the power consumption is 120W.
If you check Intel Core i9-12900K E-Cores Only Performance Review then you will find out that i9-12900K with only E-cores enabled provides 10366 points in CB R23 and consumes 118W.
You gain almost 17% higher score and power consumption is actually 2W lower.

The other interesting thing is the actual clockspeed of this i9-12900K limited to 50W.
If i9-12900K's P-cores work at 4.9GHz and E-cores work at 3.7GHz with PL1=PL2=241W and generate 27780 points in CB R23 then PL1=PL2=50W and score of 8872(or ~32% of Full performance at 241W) would mean P-cores are working at measly ~1568Mhz and E-cores at 1184MHz.

I think it's a case of the E cores being more efficient than P cores unless you drop the multiplier on the P cores to more closely match the E cores in general outside of certain workloads and instruction sets provided you don't push the E cores too high on frequency then you start to lose the efficiency advantage though overall performance will still rise. Drop the P cores frequency down to same level of E cores and they'll be more efficient. The P cores being pushed as far as they are is what makes then inefficient the same would be just as true or more with E cores being too aggressively pushed. That performance would rise with E cores being pushed higher than the P cores at the same frequencies relative to the die space of each they come out ahead.

That is why Intel has adopted this design and is expanding around increasing the number of E cores more aggressively. You can get more performance and efficiency mileage out of them over a P core though they are less specialized than a P core at certain tasks and at the same clock frequency. The P core's aren't being eliminated, but over emphasis around them is diminishing due to performance and efficiency advantages of how the E cores can be used since they pack in more performance in the same die space area.

 

[Solid State Brain]

Interesting findings. I did a similar run with my 11700, which is (theoretically) identical to your 11900 except for its different boost table and lack of TVB.

At 65 W, your 11900 scored way above 10000 points - my 11700 scored 9267.
At 125 W, yours scored around 14000 - mine scored 11700 (funny enough).
Your diagram ends a bit above 140 W and 14500 points (I guess?) - with unlocked power limits, my CPU eats around 165 W in CB R23 and scores 14100.

With this regard, it would be interesting to see how differently a 12900K and 12900 non-K limited to the same power limit perform.

Actual data was as follows in my case:

I have a good CPU and voltages have been optimized a bit (lower V at lower frequencies and higher V at higher ones—it wasn't fully stable there at default settings). I think the results strongly depend on the actual CPU tested and motherboard settings; a 12900 non-K might not necessarily be worse than 12900K, although it's likely to be.

Before the i9-11900 I had an i7-11700K, which is almost identical in terms of boost frequencies, but it didn't seem to cooperate well with my motherboard and required significantly higher voltages to be fully stable (thus it was hotter and consumed more power). I had to pay for the difference, but I think it was worth it; I might have been lucky.

 

[AusWolf]

Actual data was as follows in my case:

I have a good CPU and voltages have been optimized a bit (lower V at lower frequencies and higher V at higher ones—it wasn't fully stable there at default settings). I think the results strongly depend on the actual CPU tested and motherboard settings; a 12900 non-K might not necessarily be worse than 12900K, although it's likely to be.

Before the i9-11900 I had an i7-11700K, which is almost identical in terms of boost frequencies, but it didn't seem to cooperate well with my motherboard and required significantly higher voltages to be fully stable (thus it was hotter and consumed more power). I had to pay for the difference, but I think it was worth it; I might have been lucky.

Now I see, score = frequency for both of our CPUs - I mean, my CPU scores the same at the same frequencies. The only difference is power consumption. Your 11900 seems to be eating around 20 W less than my 11700 at lower frequencies and around 30-35 W less at higher frequencies, although I admit, I've never touched the voltages. I have a feeling that my motherboard likes supplying a bit more voltage than necessary, bumping my power consumption a bit above yours, despite the slightly lower boost table. Honestly, as long as my cooling handles it, I'm fine.

I suspect there's even more variation in voltage by motherboard choice with newer CPUs, making objective testing increasingly difficult.

 

[Solid State Brain]

It might not necessarily be eating more power in your case.

One important point for accurate Package Power readings is that the DC Load Line setting of the CPU must be set the to the same value as the VRM Load Line (i.e. what is commonly known as LLC), otherwise it might show higher or lower power. PL1 and PL2 will also work incorrectly and limit the CPU to the wrong level if it's not.






In my case, I have a "Medium" LLC which has a value of 0.7 mOhm, so I set the DC Load Line value accordingly. The AC Load Line value affects actual voltages along the entire voltage-frequency curve (when using Auto voltages), but not the power reporting.

Some motherboard manufacturers set the AC and the DC loadline setting to the same value (and the AC value tends to be low to limit voltages), but this may lead to incorrect CPU power reporting.

 

[chrcoluk]

no, that's definitely not true. Intel's efficiency curve doesn't plummet until like 150-180W.

It goes from 125w to 245w on the tests, hence my comment.

A CPU already running at 180w is not efficient in the first place in my opinion. but we all have different opinions, however to clarify, my comment is based upon TPU data.

 

[THANATOS]

I think it's a case of the E cores being more efficient than P cores unless you drop the multiplier on the P cores to more closely match the E cores in general outside of certain workloads and instruction sets provided you don't push the E cores too high on frequency then you start to lose the efficiency advantage though overall performance will still rise. Drop the P cores frequency down to same level of E cores and they'll be more efficient. The P cores being pushed as far as they are is what makes then inefficient the same would be just as true or more with E cores being too aggressively pushed. That performance would rise with E cores being pushed higher than the P cores at the same frequencies relative to the die space of each they come out ahead.

That is why Intel has adopted this design and is expanding around increasing the number of E cores more aggressively. You can get more performance and efficiency mileage out of them over a P core though they are less specialized than a P core at certain tasks and at the same clock frequency. The P core's aren't being eliminated, but over emphasis around them is diminishing due to performance and efficiency advantages of how the E cores can be used since they pack in more performance in the same die space area.

It's not about which core is more efficient at what clock. Look at this:

1.) 8* E-cores at 3.9GHz (TPU review)
Power consumption(whole system): 118 W
Performance in CB R23: 10366

2.) 8C+8c i9 12900K limited to 50W for PL1=PL2
Power consumption(whole system): 120 W
Performance in CB R23: 8872
based on this CB score the clocks should be:
8xP-cores: 1568Mhz
8xE-cores: 1184MHz

In the first case these E-cores should be clocked past their optimal point, yet It's more efficient than the second case where the clocks are extremely low. This shouldn't be happening If you think about It.

Here is a 12700K scoring 12452 in CB23 where PL1=PL2 were limited to 50W.
Anandtech forum

Then we also have scores for the mobile CPUs of the previous generation. Techspot review

i9 12900K limited to 50W scores a lot less than 11980HK at 45W. Why?
This doesn't make any sense when ADL is a new and better architecture made on a better process.
One possible explanation I can think of is that CPU voltage stayed at much higher values than needed, and with limiting power to 50W It caused low clocks.

 

[Solid State Brain]

I'll humbly suggest again to verify if clocks and voltages are correct on the 12900K using PL1=PL2="low value", because on my 11900 I have performance issues when I try to limit power in this way, as I mentioned earlier.

Setting PL2>PL1 with Tau to a very short duration (e.g. 0.5 seconds; shorter values may not work correctly) with ThrottleStop so that the processor will go right away to PL1 should be able to highlight if significant differences arise. Alternatively, PL1 could be set to a very low value, while PL2 could be set to the desired power level and Tau to infinite/very long time.

 

[AusWolf]

It's not about which core is more efficient at what clock. Look at this:

1.) 8* E-cores at 3.9GHz (TPU review)
Power consumption(whole system): 118 W
Performance in CB R23: 10366

2.) 8C+8c i9 12900K limited to 50W for PL1=PL2
Power consumption(whole system): 120 W
Performance in CB R23: 8872
based on this CB score the clocks should be:
8xP-cores: 1568Mhz
8xE-cores: 1184MHz

In the first case these E-cores should be clocked past their optimal point, yet It's more efficient than the second case where the clocks are extremely low. This shouldn't be happening If you think about It.

Here is a 12700K scoring 12452 in CB23 where PL1=PL2 were limited to 50W.
Anandtech forum

Then we also have scores for the mobile CPUs of the previous generation. Techspot review

i9 12900K limited to 50W scores a lot less than 11980HK at 45W. Why?
This doesn't make any sense when ADL is a new and better architecture made on a better process.
One possible explanation I can think of is that CPU voltage stayed at much higher values than needed, and with limiting power to 50W It caused low clocks.

Maybe work distribution and the communication between E and P cores isn't flawless when power limits are set too low.

I also don't think that E and P cores (should) change clocks in a parallel manner when PL values are modified. Pushing more work onto E cores at low PL and P cores at high PL would make more sense, imo.

 

[Melvis]

Yikes just shows how pushed this CPU is to try and match the 5950X, set at the same power limit as the 5950X and it gets destroyed.

 

[Thanamite]

With this regard, it would be interesting to see how differently a 12900K and 12900 non-K limited to the same power limit perform.

I am very interested in this for a server I am planning to build as an NVR etc. Multithreaded performance will be important but not at any power cost since the server will be up 24/7. Single threaded performance will be less important.

It looks like a PL1/PL2 setting of 75/150 would work well for me but will the same 75/150 setting give me better multithreaded performance if I wait for the non-k 12900?

Many thanks for the very interesting review.

What I find weird is that i9-12900K limited to 50W for both PL1 and PL2 provides only 8872 points in CB R23 and the power consumption is 120W.
If you check Intel Core i9-12900K E-Cores Only Performance Review then you will find out that i9-12900K with only E-cores enabled provides 10366 points in CB R23 and consumes 118W.
You gain almost 17% higher score and power consumption is actually 2W lower.

The other interesting thing is the actual clockspeed of this i9-12900K limited to 50W.
If i9-12900K's P-cores work at 4.9GHz and E-cores work at 3.7GHz with PL1=PL2=241W and generate 27780 points in CB R23 then PL1=PL2=50W and score of 8872(or ~32% of Full performance at 241W) would mean P-cores are working at measly ~1568Mhz and E-cores at 1184MHz.

These tell me that getting a 12900K and limiting its power consumption is not the same as getting a cpu optimized for the lower power consumption. This is why I am waiting for the 12900 (non-k).

 

[Solid State Brain]

In the end I got an i7-12700K. I overclocked to some extent, but I wonder how useful that is in compute multicore loads. It takes a lot of power for minimal improvements after a certain threshold. Below is a graph of render times vs processor package power that I made with Blender (note that recent versions of Blender on Linux are faster than the one tested by TPU on Windows).

The sweet spot seems to be around 105W.


Data:

Power (W)​	Render Time​
15​	06:13.00​
25​	03:29.34​
35​	02:48.59​
45​	02:25.32​
65​	01:59.07​
77​	01:52.41​
95​	01:45.16​
105​	01:41.65​
115​	01:39.49​
125​	01:38.05​
145​	01:35.23​
160​	01:33.60​
170​	01:32.51​

 

[InVasMani]

It's not about which core is more efficient at what clock. Look at this:

1.) 8* E-cores at 3.9GHz (TPU review)
Power consumption(whole system): 118 W
Performance in CB R23: 10366

2.) 8C+8c i9 12900K limited to 50W for PL1=PL2
Power consumption(whole system): 120 W
Performance in CB R23: 8872
based on this CB score the clocks should be:
8xP-cores: 1568Mhz
8xE-cores: 1184MHz

In the first case these E-cores should be clocked past their optimal point, yet It's more efficient than the second case where the clocks are extremely low. This shouldn't be happening If you think about It.

Here is a 12700K scoring 12452 in CB23 where PL1=PL2 were limited to 50W.
Anandtech forum

Then we also have scores for the mobile CPUs of the previous generation. Techspot review

i9 12900K limited to 50W scores a lot less than 11980HK at 45W. Why?
This doesn't make any sense when ADL is a new and better architecture made on a better process.
One possible explanation I can think of is that CPU voltage stayed at much higher values than needed, and with limiting power to 50W It caused low clocks.

The cache design structure of the E cores relative to die space beats the piss out of the P cores however which can't be overlooked. The L3 cache alone is double on E cores in relative terms, but it wins across the entire spectrum from a die area standpoint for every metric of L1/L2/L3 cache. If you can't figure out why that would be a positive thing you probably think 3D stacked cache is dumb as well.

 

[Xebec]

In the end I got an i7-12700K. I overclocked to some extent, but I wonder how useful that is in compute multicore loads. It takes a lot of power for minimal improvements after a certain threshold. Below is a graph of render times vs processor package power that I made with Blender (note that recent versions of Blender on Linux are faster than the one tested by TPU on Windows).

View attachment 241219
The sweet spot seems to be around 105W.

​

Very nice data and info - thank you!

Did you happen to capture the temperature at each power level? As the temperature gets hotter, the chip tends to consume more power even at the same frequency.. I'd be curious if that was part of the curve here too.

 

[Solid State Brain]

@Xebec : I didn't, but at 21 °C ambient temperature and open case (Fractal Design Define C) with Noctua NH-D15S (one fan) and moderate fan speed at the 170W power level I got 87 °C peak at 4800 MHz P-cores / 3600 MHz E-cores.

I do expect that higher temperatures will imply a higher power consumption but a detailed test to determine its exact influence would need a more powerful and controllable cooling.

 

[InVasMani]

The sweet spot balance between performance and efficiency in those charts to me appears closer to 65w. Beyond the 65w efficiency relative to performance gains really falls off quickly and steeply while the gains in performance up to that point are quite large. You save more time to render from 15w up to 65w than the entire rest of the chart from 65w up to 170w. The first 15w to 25w is enormous on efficiency. The 65w mark from a analogy standpoint is kind of your midway base camp up the mountain it's the comfortably numb spot.

 

[Solid State Brain]

At 65W, even though technically the processor works more efficiently, there's still a decent amount of performance left. I probably wouldn't cap power this low unless the PC was part of a rendering cluster operating many hours a day or even continuously. I find 105W the threshold above which it seems almost silly to put more power into the CPU for faster render speeds in this application, which is what I mostly meant.

At 65W, by the way, my CPU operates in this Blender test at about 3800 MHz (P-cores) / 3100 MHz (E-cores), hottest core at 60 °C, but I set my Noctua NH-D15S cooler to run semi-passively (mainly to avoid dust buildup) and under 62 °C its fan is not spinning.

 

[InVasMani]

Need for speed hot pursuit Intel Alder Lake Edition.

 

[JustBenching]

It's not about which core is more efficient at what clock. Look at this:

1.) 8* E-cores at 3.9GHz (TPU review)
Power consumption(whole system): 118 W
Performance in CB R23: 10366

2.) 8C+8c i9 12900K limited to 50W for PL1=PL2
Power consumption(whole system): 120 W
Performance in CB R23: 8872
based on this CB score the clocks should be:
8xP-cores: 1568Mhz
8xE-cores: 1184MHz

In the first case these E-cores should be clocked past their optimal point, yet It's more efficient than the second case where the clocks are extremely low. This shouldn't be happening If you think about It.

Here is a 12700K scoring 12452 in CB23 where PL1=PL2 were limited to 50W.
Anandtech forum

Then we also have scores for the mobile CPUs of the previous generation. Techspot review

i9 12900K limited to 50W scores a lot less than 11980HK at 45W. Why?
This doesn't make any sense when ADL is a new and better architecture made on a better process.
One possible explanation I can think of is that CPU voltage stayed at much higher values than needed, and with limiting power to 50W It caused low clocks.

Yeah, the tests are wrong. 12900k at 75w score 18500-19200, and at 125w it score over 23500. We tested 3 different CPUS at 4 different motherboards. There is something very weird going on with this review

Yikes just shows how pushed this CPU is to try and match the 5950X, set at the same power limit as the 5950X and it gets destroyed.

No it doesn't, it's just that the numbers on the review are plain wrong.

@TheoneandonlyMrK Are you laughing cause you know im right? It's not just me saying this, there are other users on other forums, like this from anandtech

Page 59 - Question - Alder Lake - Official Thread

Page 59 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com