Question Anandtech.com article on 13900k and 7950x power scaling

[adamge]

A Lighter Touch: Exploring CPU Power Scaling On Core i9-13900K and Ryzen 9 7950X

www.anandtech.com

It's an interesting article, but IMO severely crippled by the reliance on "configured" power limit rather than measuring power usage. It's even more frustrating when one of the datapoints is measured power usage, so the author has access to this data.

In this scenario where configured power limits are basically hand waving guidelines, measured power usage (specifically, average power usage by the CPU during the benchmark run) is the only useful, meaningful, and concrete way to analyze the performance results relative to the power usage.

 

[TESKATLIPOKA]

Nah, we have forum users hitting better scores than that 38k on their 7950X @ below 200W.

I meant based on temperature, It would be
65W 7950X 52°C vs 125W 13900K 53°C
CB R23: 31179(100%) vs 31615(101%)

105W 7950X 78°C vs 125W 13900K 53°C
CB R23: 35975(114%) vs 31615(100%)
These temperatures are with a water cooler.
To me It looks like 7945HX will be held back by temperature to pull ahead of i9 13980HX in CB R23 for example.

 

[TESKATLIPOKA]

That's the same dies, but very likely not the same behavior set in the firmware. E.g. STAPM (Skin Temperature Aware Power Management) is very likely a major limiting factor in Dragon Range that wasn't in Raphael.

What does It do exactly? When a temperature sensor senses a high temperature, It will lower the power limit, right?
Then performance will go down.

 

[TESKATLIPOKA]

These graphs don't show clocks.
I think Zen 4 uses much less power to hit a certain clock and performance, it's just that when a certain clock is reached the power consumption and temperature rises disproportionally. Lower the clock a bit and the temperature problem is solved.
Also, mobile APUs are monolithic.

Didn't AMD made exceptional claims about the battery autonomy?

As @Thunder 57 already mentioned, there are 2 different laptop AMD CPUs(APUs).
One is called Dragon Range and that is desktop Raphael, this one is a chiplet design N5 + N6 process.
Then there is Phoenix which uses N4 process, 8C16T with RDNA3 IGP, AI chip.

 

[TESKATLIPOKA]

What I am saying is you can't infer much from that. Until the mobile parts are in the wild and actually benchmarked then it's speculation.

It's the same chip, probably better binned. You can't expect miracles from that.
If 7950x 105W with a water cooler manages only 78°C then a laptop 7945HX + 150W GPU with a shared air cooling solution shouldn't be able to cool It at 105w power limit. That’s my opinion, but maybe I am just a pessimist. Many people told me that.

 

[TESKATLIPOKA]

This is also interesting

The best and second best score for 6900HX was 14,584-15,100 points at notebookcheck.
7945HX should score 25,960-26,878 points.
That's a lot less than what 7950X manages at 65W TDP(88W PPT).
They probably limited It to ~55W.

 

[Timur Born]

As I wrote in the comments section for the article I'm grateful for any new content on Anandtech. But with that being said...

1. More power usage data points needed to draw conclusions. 125W and then a jump to 230W. Too large. There is a lot going on in that gap that needs to be investigated.

I agree, especially the 230W and 253W points should have been used for both CPUs for direct comparison.

2. Benchmarks aren't the be all end all application for power usage. Outside of benchmarks not a lot of applications use all of the cores or even use all of the resources in a core. So while 5.5GHz all cores loaded under Cinebench might result in a ridiculous power usage for the 13900K, in a real world application such as an NLE those frequencies are reached an maintained with a package power of 160W in many cases. This can be seen in the gaming benches, which are real world and things even out quite a bit.

While I agree on the main point I disagree that Cinebench is not a real world application, it essentially represents Cinema 4D performance. One point I would like to know more about in this regard is low/medium load applications or even games. My 5900X used considerably more power to run WoW than my 9900K using the same settings and running at the same 60 fps via the same GPU.

3. Bursty loads aren't taken into account either and it should be noted that cranking power up to 250W for 3 seconds every now and then while working so you aren't waiting on the computer might be a worthwhile tradeoff in terms of keeping a creative flow going.

The 13900K does brute-force its way up to 7950X competitive multi-core performance, there is no doubt about that. But I still decided to go Intel this time around for the single/low core performance and power usage.

But I am really disappointed at Intel that I was able to lower my CB23 power usage by as much as 20% (twenty!) via undervolting. This is far too much headroom/tolerance for a 750 EUR CPU (now only 630 EUR). It should have arrived using much tighter stock settings out of the box, especially since CPUs nowadays already crossed the point of diminishing returns long before leaving the factory.

 

[moinmoin]

What does It do exactly? When a temperature sensor senses a high temperature, It will lower the power limit, right?
Then performance will go down.

Essentially this. It's supposed to take a chassis temperature close to the user's skin and keep temperature there in check to not become burning hot. Which means the way the cooling system is implemented and where exactly that temperature sensor is located influences everything.

They probably limited It to ~55W.

Which is likely why using PBO will allow for a significant performance boost on top.

 

[Mopetar]

I agree, especially the 230W and 253W points should have been used for both CPUs for direct comparison.

Zen 4 caps out at 230W so it wouldn't be possible to have a 253W data point. The difference between the two for an Intel chip is probably pretty negligible as well.

I'm not sure how useful the comparison would be since it's clear that Zen 4 scales poorly past 105W. You could even say that it doesn't scale particularly well past 65W since if you look at the Cinebench results going from 65W to 125W only gives an 18% performance bump. Intel gets a much more modest 38% boost in the same range, but it basically needs it to have a comparable score to AMD.

 

[IEC]

It seems like posters are getting confused by how complex the clocking behavior of modern CPUs is, so let me help clarify.

0) Both the Ryzen CPUs and Raptor Lake CPUs are factory OC'd to the limits (beyond sane limits).

When in doubt, refer to fact #0 above

1) voltage required for a given frequency scales exponentially the higher you go up the f curve
2) power required also scales exponentially with voltage (simple physics)
3) the CPUs are manufactured on different processes and have different strengths and weaknesses
4) there's a point on the v/f curve where the increased voltage (and consequentially, power usage) leads to miniscule gains in frequency - with mild variance between individual samples
5) these CPUs have so many sensors on them that there are MANY reasons why performance can throttle - temperature, power limit, voltage, etc.
6) the SI unit that should be used to measure power used to complete a task is Joules (J), which I've only seen used in CPU reviews a handful of times (probably because of the difficulty of isolating CPU power usage)
7) when measured in Joules, I suspect Zen 4 easily takes the efficiency cake, even more so if set to a sane power limit (e.g. 120W TDP with the upcoming X3D chips). At stock 170W TDP it burns power for minimal (or in cases of poor cooling, negative) gains
8) TDP != PPT

This is how you end up being able to cap a Ryzen 7700X at 95W PPT (again, PPT != TDP) and match stock performance while using significantly less power versus the stock 105W TDP/142W PPT. My chip's personal "sweet spot" which isn't significantly different from observed behavior with Zen 3.

I don't think it's a coincidence AMD chose 120W vs 170W for their X3D lineup. These chips would have received rave efficiency reviews while losing low single digit % performance at 120W TDP (2 chiplet models). They are made on a process that has superior power efficiency characteristics versus their competitor. But AMD tried to squeeze out the last few % since they knew they had the more efficient design and could get away with it.

This is somewhat confounded at the low end of TDPs (e.g. 35W) due to the chiplet design requiring some efficiency compromises, but that's why a dedicated mobile chip (e.g. 7040) will always be superior in that regard.

tl;dr limit your PPT for Zen 4 to ~70% of stock settings and lose little to no performance for massive power savings

 

[TESKATLIPOKA]

It seems like posters are getting confused by how complex the clocking behavior of modern CPUs is, so let me help clarify.

0) Both the Ryzen CPUs and Raptor Lake CPUs are factory OC'd to the limits (beyond sane limits).

When in doubt, refer to fact #0 above

1) voltage required for a given frequency scales exponentially the higher you go up the f curve
2) power required also scales exponentially with voltage (simple physics)
3) the CPUs are manufactured on different processes and have different strengths and weaknesses
4) there's a point on the v/f curve where the increased voltage (and consequentially, power usage) leads to miniscule gains in frequency - with mild variance between individual samples
5) these CPUs have so many sensors on them that there are MANY reasons why performance can throttle - temperature, power limit, voltage, etc.
6) the SI unit that should be used to measure power used to complete a task is Joules (J), which I've only seen used in CPU reviews a handful of times (probably because of the difficulty of isolating CPU power usage)
7) when measured in Joules, I suspect Zen 4 easily takes the efficiency cake, even more so if set to a sane power limit (e.g. 120W TDP with the upcoming X3D chips). At stock 170W TDP it burns power for minimal (or in cases of poor cooling, negative) gains
8) TDP != PPT

This is how you end up being able to cap a Ryzen 7700X at 95W PPT (again, PPT != TDP) and match stock performance while using significantly less power versus the stock 105W TDP/142W PPT. My chip's personal "sweet spot" which isn't significantly different from observed behavior with Zen 3.

I don't think it's a coincidence AMD chose 120W vs 170W for their X3D lineup. These chips would have received rave efficiency reviews while losing low single digit % performance at 120W TDP (2 chiplet models). They are made on a process that has superior power efficiency characteristics versus their competitor. But AMD tried to squeeze out the last few % since they knew they had the more efficient design and could get away with it.

This is somewhat confounded at the low end of TDPs (e.g. 35W) due to the chiplet design requiring some efficiency compromises, but that's why a dedicated mobile chip (e.g. 7040) will always be superior in that regard.

tl;dr limit your PPT for Zen 4 to ~70% of stock settings and lose little to no performance for massive power savings

I really like this post of yours. Great job writing It.
I must point out that with R7 7800x3D AMD actually increased TDP from 105W to 120W, but boost was lowered from 5.4GHz -> 5GHz.

I don't think the increase in TDP would bring noticeable performance gains. I would have expected a lower TDP because of only a single CCD with that extra cache.

 

[Hulk]

I agree, especially the 230W and 253W points should have been used for both CPUs for direct comparison.


While I agree on the main point I disagree that Cinebench is not a real world application, it essentially represents Cinema 4D performance. One point I would like to know more about in this regard is low/medium load applications or even games. My 5900X used considerably more power to run WoW than my 9900K using the same settings and running at the same 60 fps via the same GPU.


The 13900K does brute-force its way up to 7950X competitive multi-core performance, there is no doubt about that. But I still decided to go Intel this time around for the single/low core performance and power usage.

But I am really disappointed at Intel that I was able to lower my CB23 power usage by as much as 20% (twenty!) via undervolting. This is far too much headroom/tolerance for a 750 EUR CPU (now only 630 EUR). It should have arrived using much tighter stock settings out of the box, especially since CPUs nowadays already crossed the point of diminishing returns long before leaving the factory.

But I am really disappointed at Intel that I was able to lower my CB23 power usage by as much as 20% (twenty!) via undervolting. This is far too much headroom/tolerance for a 750 EUR CPU (now only 630 EUR). It should have arrived using much tighter stock settings out of the box, especially since CPUs nowadays already crossed the point of diminishing returns long before leaving the factory.

This is interesting and I think worthy of further discussion. If Intel tests and creates a specific v/f curve, which is built into the CPU and read by the motherboard, why do some people report being able to undervolt their CPU's to such a large degree?

Of course I don't know but have some ideas.

First, as we know these parts are binned and some are better than others. With a mere -0.025 undervolt I've had my 13900K pull a restart under heavy load when drawing 250W. Cooling not good enough? Perhaps but I have a 280mm radiator, which is currently out of the case. My point is that I think it is possible that some undervolts are not as stable as they seem. I can go -0.1V if I keep the power limit to 175W. So the amount of undervolt is not only determined by the bin quality but I think also by cooling/watts. Furthermore, there is only so much heat that can be moved from the cpu to the cooler so it's not so much the radiator that is the problem but the cold plate. Hot spots I think begin to crop up as failure points. To sum up point #1, Intel has to make sure their v/f curve works at all power levels up to and most likely beyond 253W. Does your undervolt hold up at 253W running CB 23 and Prime 95 for half hour each?

Second. Intel doesn't know the quality of the motherboard you will be using and has to assume it's not great from a current and voltage stability point of view. Their v/f curve is set for low end boards.

Third. Intel doesn't know about your cooling. Will you be running your chip at 90C most of the time under load or at 70C? I bet the volts needed are different.

Fourth. When comparing various 13900K's at different power levels we don't know how much of the fact that your CPU uses less watts than mine at a certain performance level is your undervolt or the fact that you have a better binned part or a combination of both. So when someone says/writes they can hit this score with this much watts because of their undervolt, say 30 watts better than someone else. 20W of that difference might be simply because they have a better CPU with a better v/f curve and the other 10W undervolting, which might fail at higher watts than they have set in their BIOS.

Yes, it's complex.

 

[DrMrLordX]

I can imagine certain scenarios where Raptor Lake's efficiency shines, owing to the E cores. It's just that its P cores are very power hungry. For mixed use scenarios, Raphael and Raptor Lake are a wash for both performance and power use. (imo of course)

That's only cuz Raptor Cove cores are clocked too high. Clock them down a little and they catch up to Gracemont in power efficiency. It's area where Raptor Cove is inefficient.

I meant based on temperature

How is that comparison relevant? If Zen4 is using less power to reach the same level of performance as Raptor Lake, who cares what is the temperature of either CPU at that point?

 

[TESKATLIPOKA]

How is that comparison relevant? If Zen4 is using less power to reach the same level of performance as Raptor Lake, who cares what is the temperature of either CPU at that point?

I am talking about when they are using the same power. AMD should be faster in CB R23, but temperature can't be too high, because It will start throttling.

Temperature is too high because of insufficient cooling resulting in power draw being limited and because of that clocks are also reduced.




If a laptop cooling can keep temperature to a level comparable to Wraith cooler at 80% fan speed then 7945HX(Dragon Range) could be set to 105W(142W PPT) and will be ~14% faster than 13980HX at similar power draw. If It can't then the difference in performance will be smaller, at worst It will be similar but at ~1/2 of power draw.

P.S. Regrettable that TPU didn't also include power draw for Wraith at 20% or achieved scores in CB R23.

 

[DrMrLordX]

but temperature can't be too high, because It will start throttling.

That's not a terribly compelling argument when none of the temperatures you've listed cause throttling. Lower temps CAN help Zen4 reach better efficiency points on its temp/voltage/clockspeed curve, but it's still more-efficient than Raptor Lake @ 90C.

 

[Timur Born]

To sum up point #1, Intel has to make sure their v/f curve works at all power levels up to and most likely beyond 253W. Does your undervolt hold up at 253W running CB 23 and Prime 95 for half hour each?

I specifically use power limits to keep Prime 95 load stable. My CB 23 load does not reach 253W, but is stable at around 240W. The more voltage I use the higher I can push the power limit and stay P95 stable, but the idea is to lower realworld load and keep it below the power limit, so that would be counterproductive.

The undervolted (and maybe per core overclocked) CPU must be as stable as a stock one up to the power limit, regardless of what you throw at it. This is mandatory for me as I use my PC for both leisure (mostly *not* fps gaming) and work.

Second. Intel doesn't know the quality of the motherboard you will be using and has to assume it's not great from a current and voltage stability point of view. Their v/f curve is set for low end boards.

They could dictate less lose tolerances for board manufacturers. And in any case, 20% is just too much headroom.

Third. Intel doesn't know about your cooling. Will you be running your chip at 90C most of the time under load or at 70C? I bet the volts needed are different.

Yes, but then my fans (and coupled pump) run at 400 rpm most of the time, do not max out most of the time and only max out at 1600 rpm when they do. 20% still seems too much of a headroom.

Fourth. When comparing various 13900K's at different power levels we don't know how much of the fact that your CPU uses less watts than mine at a certain performance level is your undervolt or the fact that you have a better binned part or a combination of both.

For 750 EUR (now 630 EUR) everyone should have gotten a proper bin with tighter stocks settings then.

 

[Hulk]

I specifically use power limits to keep Prime 95 load stable. My CB 23 load does not reach 253W, but is stable at around 240W. The more voltage I use the higher I can push the power limit and stay P95 stable, but the idea is to lower realworld load and keep it below the power limit, so that would be counterproductive.

The undervolted (and maybe per core overclocked) CPU must be as stable as a stock one up to the power limit, regardless of what you throw at it. This is mandatory for me as I use my PC for both leisure (mostly *not* fps gaming) and work.


They could dictate less lose tolerances for board manufacturers. And in any case, 20% is just too much headroom.


Yes, but then my fans (and coupled pump) run at 400 rpm most of the time, do not max out most of the time and only max out at 1600 rpm when they do. 20% still seems too much of a headroom.


For 750 EUR (now 630 EUR) everyone should have gotten a proper bin with tighter stocks settings then.

You win. Contact Intel and tell them your ideas ASAP!
Just kidding. You quote 20% but I'm not seeing that with my particular chip due to the reasons I listed. By your standards I'd be returning my chip to Intel as defective. And that's what they are avoiding by being conservative

 

[PJVol]

0) Both the Ryzen CPUs and Raptor Lake CPUs are factory OC'd to the limits (beyond sane limits).

Not quite the right term - "OC'd", since these CPUs never intended to run at base clocks.

1) voltage required for a given frequency scales exponentially the higher you go up the f curve

No )

4) there's a point on the v/f curve where the increased voltage (and consequentially, power usage) leads to miniscule gains in frequency

You mean minuscule above ~ 7-8 gHz ?

(mostly agree on the rest)

 

[Khanan]

Whoever argues in favor of the 13900K Being more efficient than the 7950X is delusional. No amount of tweaking will make the 13900K be more efficient than a similar tweaked 7950X

Of course it is, this is 5nm vs 10 (or 7 whatever they call it now), and Ryzen is well known for being more efficient than intel. The last 2 gens of intel CPUs are known for pushing way past the sweet spot to be barely competitive with AMD, the 12900K was a bit ridiculous, at least the 13900K didn’t increase the numbers much, so it’s clearly better but not better than the 7950X.

Suffice to say it has a reason why AMD has the most efficient (and fastest) servers on the planet not intel. The same chips are used in 7950X with a different configuration.

 

[Timur Born]

You win. Contact Intel and tell them your ideas ASAP!
Just kidding. You quote 20% but I'm not seeing that with my particular chip due to the reasons I listed. By your standards I'd be returning my chip to Intel as defective. And that's what they are avoiding by being conservative

I can do per core overclocking at my current settings, so there is even more headroom. As a consequence I will lower the power limit to see how low you can go for realworld load. Currently I am peaking 232W in CB23, which would be 25% down from stock. But this would have to be properly stability tested (and I test more than any OCCT Platinum test could) and I seem to hit a point where my BIOS won't let me go lower than this (combination of AC/DC + LLC + offset keep resulting in the same low point).

 

[TESKATLIPOKA]

That's not a terribly compelling argument when none of the temperatures you've listed cause throttling.

Those temperatures were achieved by a high-end water cooler.
Laptops aren't cooled by water, except a few special cases.

That water cooler needed to cool 230W, still the achieved temperature was 94 °C.
That's only 1 °C lower than what AMD has set as a target for temperature to be achieved.
Temperature can't go over this value, and Zen 4 will boost clocks and voltage until It reaches this target temperature during load or power limit.

Anand had a powerful cooler and set different power limits, so It's not surprising temperatures didn't get close to 95 °C unless It was 170W(230W PPT).
TPU did a test about Cooling Requirements & Thermal Throttling for 7950X. Link
A very good read. The last few charts were taken from there.

Only that Arctic AIO is not hitting 95 °C. The rest is hitting It, which causes a reduction in clocks and power draw.

Arctic AIO - frequency: 5240 MHz(100%); power draw: 230W (100%)
NH-U14S 100% - frequency: 5171 MHz(98.5%); power draw: ~215W (93.5%)
NH-U14S 20% - frequency: 4718 MHz(90%); power draw: ~170W (74%)
Wraith Spire 100% - frequency: 4660 MHz(89%); power draw: 150W (65%)
Wraith Spire 40% - frequency: 3839 MHz(73%); power draw: ~102W (44%)
Wraith Spire 20% - frequency: 3231 MHz(61.5%); power draw: 65W or less?
As shown, the weaker the cooler is, the more throttling we see.
We are talking only about cooling a CPU, what about If you included 150-175W GPU, which will happen in a laptop?

What's my conclusion?
7945HX(Dragon Range) will be faster than 13980HX(Raptor Lake), If manufacturers don't limit It to low power limits.
Why the change in opinion? I just realized after checking different laptop reviews, that CPU is not much used during gaming, or that GPU is not used during rendering.
It could be a problem with stress test like Prime95 + FurMark which notebookcheck does, but that's not important.

TPU review also showed that during Cyberpunk gameplay 7950X stock had only 76 °C and that was with air cooler NH-U14S, not sure at what Fan speed.

It shouldn't be a problem even If you pair 125-150W 7945HX with 175W RTX 4090 Laptop.
High-end laptops have cooling capable of ~250W and during gaming even If GPU is consuming 175W, CPU is likely consuming less than 75W.

Lower temps CAN help Zen4 reach better efficiency points on its temp/voltage/clockspeed curve, but it's still more-efficient than Raptor Lake @ 90C.

Lower temps don't help efficiency, what helps efficiency is If you either set power limit to a lower value or cooling system is so weak that temperature is hitting target temperature which is 95 °C for stock 7950X and causes throttling.

 

[moinmoin]

That water cooler needed to cool 230W, still the achieved temperature was 94 °C.
That's only 1 °C lower than what AMD has set as a target for temperature to be achieved.

That only means that in that case another limit was hit that keeps temperature from reaching 95°C. That's Raphael's intended behavior. It should be natural that a mobile chip like Dragon Range shouldn't and won't follow that same behavior.

 

[TESKATLIPOKA]

That only means that in that case another limit was hit that keeps temperature from reaching 95°C. That's Raphael's intended behavior. It should be natural that a mobile chip like Dragon Range shouldn't and won't follow that same behavior.

I know. Power limit was hit.
I wonder If the difference between Dragon Range and Raphael won't be only in setting temperature target to 85-90 °C and STAPM (Skin Temperature Aware Power Management).

 

[Khanan]

The heat spreader of Ryzen 7000 isn’t good and that’s why they want these chips to run as hot as possible to be still able to hit certain marks.

 

[Hulk]

I can do per core overclocking at my current settings, so there is even more headroom. As a consequence I will lower the power limit to see how low you can go for realworld load. Currently I am peaking 232W in CB23, which would be 25% down from stock. But this would have to be properly stability tested (and I test more than any OCCT Platinum test could) and I seem to hit a point where my BIOS won't let me go lower than this (combination of AC/DC + LLC + offset keep resulting in the same low point).

Can you provide an overview of your BIOS settings? Obviously you have it tweaked nicely.
Also does your mobo provide the SP rating of your CPU?
Finally what is your PL1/PL2 setting and what is your CB R23 score at that setting?

I have a "hot" core 5 so I might be able to do better if I cap that one at a lower frequency.

 

[Timur Born]

Lower temps don't help efficiency, what helps efficiency is If you either set power limit to a lower value or cooling system is so weak that temperature is hitting target temperature which is 95 °C for stock 7950X and causes throttling.

I find this rather surprising as I see package power usage increasing slightly the higher the temperature gets with the 13900K. And with my MSI X570 (5900X) the hotter the VRMs got the higher system power usage increased under the same load (again within reason).