Intel Core i9-14900K Raptor Lake Tested at Power Limits Down to 35 W

[Imouto]

No, but LOL you are so predictable. Please at least try to understand some details for once:

Dr. Dro is talking about a 5950X which uses a 12nm IO die and about using 10W, while Anandtech tested a 7950X which uses a much lower power 6nm IO die and at ~35W handily beat the 13900K in efficiency.

It doesn't matter that the 7950X beats the 13900K all the way from stock to 46W which is miles away from any production/media environment and I'd say even from a scholar standpoint. Because it wins at that single extremely niche data point we have to hear of it in the comments of every single article about CPU efficiency.

You can almost move up AMD in every tier in Anand's testing and AMD would still come ahead. But 35W and 35W alone is what matters.

You can't win against fevgatos unless you outlast him. So take that in mind when you think you are ready to move on.

 

[JustBenching]

It doesn't matter that the 7950X beats the 13900K all the way from stock to 46W which is miles away from any production/media environment and I'd say even from a scholar standpoint. Because it wins at that single extremely niche data point we have to hear of it in the comments of every single article about CPU efficiency.

You can almost move up AMD in every tier in Anand's testing and AMD would still come ahead. But 35W and 35W alone is what matters.

You can't win against fevgatos unless you outlast him. So take that in mind when you think you are ready to move on.

Nice strawman though, never said anything remotely close to what you are saying. You made it up

 

[Imouto]

Nice strawman though, never said anything remotely close to what you are saying. You made it up

It's the fastest CPU at 35w, hence the most efficient. So - what do you mean "whats the point"?

The only thing that matters to you is "the most efficient" title no matter how pointless is how it's achieved.

 

[JustBenching]

The only thing that matters to you is "the most efficient" title no matter how pointless is how it's achieved.

I stated a fact. Never said it's important or anything else isn't important.

 

[BoggledBeagle]

Anyway, in my opinion 125 W should be the default turbo power limit of all these Intel K CPUs.

Nope, not at all. After a lot of playing with 13900K at different power limits, I can tell you that the right sensible power limit is somewhere between 150-180W. Trust me.

125W for 24 core cpu is simply too low. Too bad that W1zzard did not test anything between 125 and 253W, it is a huge gap.

BTW power draw limit is not the only limit that should be changed - 6 GHz is simply too much for the Intel process, 5 GHz is just right.

5 GHz 160W 14900K is such a nice cool efficient CPU, unfortunately a little bit slower than AMD offerings, which are hitting thermal limits... Is it a real problem or not? Would Intel sell considerably lower number of CPUs, if they were sold with sensible settings and they would not be in the first spot in review graphs???

 

[Gica]

The notion of "E-cores being detrimental" has to do with Alder Lake processors having the ring bus clock tied to the E-core top frequency. In Raptor Lake, this does not apply: the ring and E-core clock domains are completely decoupled, and with the increases in E-core speed, they are comparable in speed to a Sky or Kaby Lake core, but with an impossibly small footprint. Not too bad!

There were also software problems, inherent in the beginning, both from Intel, but also Microsoft and many others had to work with the settings.
Now the processor cores are completely independent and the software is matured. If any developer still stutters, the joy of E-cores detractors. Even if the first update solves the problem, the damage has been done and the dwarves must be executed. AMD had much more serious problems (some very serious, with metal melts) but it doesn't matter - E-cores must be annihilated.

Another crazy dance observed on all forums is that of games. The processor would matter a lot. Hmmm, I don't think so.
3070 Ti ~= 2080 Ti, so all owners of RTX 2000 or lower, as well as a good part of those who own 3000/4000 or 6000/7000 will not get miracles from a powerful processor. Probably over 80% of players.

 

[THU31]

@Gica

You have said many sensible things in your arguments, but using 3DMark as an example of CPU gaming performance is extremely bad.

The graphics tests have always been 99% GPU limited, that's the whole point of these tests. That's why there's a separate CPU benchmark, which is also heavily multi-threaded, so the CPU score is kind of useless from a gaming perspective, just like the overall score.

This year we have had many CPU-limited games, mainly Unreal Engine 4 titles, and pretty much all of those games were not able to utilize more than 6 cores, some of them even just 4 cores. Having a CPU with the highest IPC was extremely beneficial in those titles. Older AMD and Intel CPUs would cause those games to drop to ~40 FPS no matter what GPU you had.

 

[Gica]

I also tested in games and said that I don't see any difference between 13500 and 14700KF. I think like on the 14700K topic. I didn't even expect it because between 10500 and 13500 there were tiny differences in the same games. Games that I have and that I use, some occasionally. Let's agree that no one can play all the titles available on the market.

Another test: 14700KF PL1/2: 150W versus 250W. Apparently, 150W wins, but I suspect the intervention of other programs during the running of the single test where the differences are large. I have all the programs in the background, I haven't even removed Afterburner and HWinfo.

 

[Akkedie]

This is really superb work W1zzard, bravo! Answered every question I had.

 

[Gica]

Because of that huge update (67GB) I left this masterpiece of optimization disaster, Cyberpunk, at the end.
According to the TPU review, the 14700K eats ~170W in game. I'm going to tell you a big secret: it consumes even more if the image is static (such as the display of results) or if you drag the game "into the bar". If you want a break, it is better to leave the game active, not paused or sent to the taskbar.
It is one of the games that raises the consumption average in reviews to the sky.
During the benchmark run, I estimate the consumption at ~125W with the stock processor and <50W with the same settings but with the Core iccMax set from unlimited (stock) to 170A. The results (Min, Max, Average) are identical, within the margin of error. That's why, for the lazy or ignorant, a non-K processor is recommended if they want Intel. The K variants are left free and get to use the cannon to kill flies.

 

[THU31]

@Gica

Hmm. Could you say a bit more about the IccMax option, what it does exactly and how it works?

In this situation it basically reduced the P-core clock from 5.5 to 4.4 GHz (which drastically lowered the voltage)? You were GPU limited, so it didn't really affect the result, yes?

What would happen in a CPU-limited scenario? Let's say a game that can only use 4 cores, or if you swapped the GPU to a 4090 with other variables unchanged.

 

[Solid State Brain]

That's why, for the lazy or ignorant, a non-K processor is recommended if they want Intel. The K variants are left free and get to use the cannon to kill flies.

Don't the non-K variants currently also have limitations in term of voltage / loadline control, which would be of interest to those who want to undervolt their CPUs? Or at least, last time (13th gen) there were some in place.

Hmm. Could you say a bit more about the IccMax option, what it does exactly and how it works?

IccMax affects the maximum allowed instantaneous current inside the CPU; it has a higher effect with certain workloads.

 

[W1zzard]

If just 1 then 200W, if 2 then 165W and 210W. Thanks!

Test results for 200 W have been added

 

[BoggledBeagle]

I did a test with 14700K at 25W increments.

Sensible limit for 14700K is 150W IMO and 14900K 170-180W, it has just 4 small e-cores more.

 

[FrostWolf]

@W1zzard Thanks for running the 200w tests. This indicates to me at least that one can get much more sane temperatures and lower power while still close to full performance. For the moment, I plan to lock my system there once everything is back together.

 

[AnotherReader]

Looking at the results again, I'll repeat an obvious point: not all applications respond equally to a lower power limit. Applications that make heavy use of AVX, such as Blender, show a significant regression at power limits lower than or equal to 125 W. Notice that the 7950X3D is significantly faster despite consuming only 15W more and paying the penalty of the IO die and the off chip interconnect.

 

[THU31]

Applications that make heavy use of AVX, such as Blender, show a significant regression at power limits lower than or equal to 125 W. Notice that the 7950X3D is significantly faster despite consuming only 15W more and paying the penalty of the IO die and the off chip interconnect.

When me and my friend finished building his new PC with a 7800X3D, we used Prime95 to test temperatures. Prime95 with AVX resulted in lower power draw and temperatures than without AVX (clock speeds were pretty much the same).

I found that interesting, as on Intel, Prime95 with AVX uses drastically more power.

 

[AnotherReader]

When me and my friend finished building his new PC with a 7800X3D, we used Prime95 to test temperatures. Prime95 with AVX resulted in lower power draw and temperatures than without AVX (clock speeds were pretty much the same).

I found that interesting, as on Intel, Prime95 with AVX uses drastically more power.

That's rather unexpected. AVX-512 should use slightly less power than AVX2 for Zen 4 because of less front-end resource consumption, but it should still consume more power than SSE.

 

[Gica]

@Gica

Hmm. Could you say a bit more about the IccMax option, what it does exactly and how it works?

In this situation it basically reduced the P-core clock from 5.5 to 4.4 GHz (which drastically lowered the voltage)? You were GPU limited, so it didn't really affect the result, yes?

What would happen in a CPU-limited scenario? Let's say a game that can only use 4 cores, or if you swapped the GPU to a 4090 with other variables unchanged.

The current limit drawn by the CPU.
With a shortage of amperage, it seems that the software treats the processor as a non-K.

As I have said so many times, I did not expect performance increases by replacing 13500 with 14700KF. The i5-13500 has a reserve for the 4070Ti and I rely on the fact that the processor "dances" between 30 and 60W in all games with the 3070Ti. Including in Cyberpunk. It exceeds the 60W threshold only when loading scenes.
It is very likely that 170A will be insufficient for the RTX 4090, but it offers too much for weaker video cards than the RTX 3070Ti and in no case are the 180W seen in the TPU review justified. And it's not a mistake because, left free, the 14700K draws even 145W with the 3070Ti.

The current limit of 170A is equivalent to PL 105W. In the case of games optimized for four cores, as you say, most of this power will be directed towards them. And it is enough.
As you can see, in the XTU benchmark, the CPU came very close to the stock frequency with the wattage limitation at 105W. According to this benchmark, it loses 17.5% of performance using 37.5% of power. I am convinced that there is no difference in single-threaded applications or those that can handle 105W.
Anyway, everyone can set the correct values according to their configuration.

Don't the non-K variants currently also have limitations in term of voltage / loadline control, which would be of interest to those who want to undervolt their CPUs? Or at least, last time (13th gen) there were some in place.

You don't really need to adjust the efficiency of the non-k processors. The software does an excellent job.

 

[AleXXX666]

1. what is the point of testing such a CPU in 65W mode? just for use case, like some loser have got this stolen on some hot sale and put into looser s*ckface H610 motherboard, and OEM also maybe? Or some genius using this with Intel Box Cooler?
2. How on earth i5-10400F is FASTER than i5-11400F, when in real world life i5-11400F is comparable to i7-10700K@stock ? At least in games...

 

[JustBenching]

Test results for 200 W have been added

What's up with your ycruncher numbers? They seem too low for all CPUs. I know it's memory sensitive but not that much, I get 58s with a 12900k and here you have the 7950x at over 60.

Are you running 5b and just the title is wrong?

 

[tommys]

We used joules in the previous cpu bench and nobody understood it, so i changed it to a more familiar way of saying the same thing by exposing watts instead of joules

What about Watt seconds? It's equal to Joule and is self-explanatory

 

[chrcoluk]

On Raptor Lake is very important what limit will be used - Current limit (A) or PL (W). In my experience the Current limit(A) gives more power efficiency than the Power Limit(W).

Any suggested values based on your experience?

I did a test with 14700K at 25W increments.

View attachment 319628

Sensible limit for 14700K is 150W IMO and 14900K 170-180W, it has just 4 small e-cores more.

Thanks for the data, I mostly agree with you, I think 100 - 150w is best options range on your data, 150w if you dont mind some loss of efficiency whilst not over crippling the CPU. Although silicon lottery plays a part of course, I remember getting a bigger score drop off than you at 125w on my 13700k, but with undervolt it improved the situation considerably allowing higher clocks at that power limit.

I did lower my power limit to 175w due to poor case air flow a while back, was meant to raise it again after improving it but never got round to it, but I think now I dont care so much about peak multi threaded performance, we already know low threaded workloads are mostly not impacted with lower PL which pretty much covers the majority of software I use particularly games. I dont software encode anymore, but if I ever do start using my main desktop CPU for compiling, encoding whatever I will prefer efficiency over peak performance on that type of workload so lower PL it is moving forward.

--

Also on the points made in this thread, anything related to watts/points watts/second is how I personally prefer the data to be presented, it is far easier to understand and take in than just showing the watts and a score.

 

[maukkae]

What happened to the 95W undervolt results?

 

[chrcoluk]

What happened to the 95W undervolt results?

They got pulled?

Yeah they gone.

@W1zzard