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.

 

[In2Photos]

That's because you're old school. Most review sites have YouTube channels these days (and the ones that don't are likely floundering), and some of these YouTube channels are better than the older review sites like Anandtech.

I shouldn't need to say that Anandtech has been going downhill for a long time now, and YouTube channels like Linus Tech Tips, Gamers Nexus, Hardware Unboxed etcetera are blowing up.

der8auer is a great technical source for people that are into tweaking and overclocking and hardware modifications and he's very competent.

This may mean that hell is freezing over but I have to agree with Carfax here, der8auer is one of the more competent, experienced, and well respected YouTubers in the tech industry. Roman is the person behind many of the recent advancements in cases thanks to the O11 series and has also worked with Thermal Grizzly on several products.

 

[Hulk]

Raptor Lake is a good processor but it is also desperately in need of the Intel 4 process. If they could cut power by 20% it would be fantastic. But as we know when processes shrink heat transfer and hot spots become an issue.

In order to mitigate hot spots I wonder if interspersing the P's and E's would help? Or would that impact intra-chip communications? Or even with current Zen 4 high core could parts if the Thread Director could know that say only 8 out of 16 cores are heavily loaded then activate 4 cores on each chiplet at high speeds and 4 at lower speeds to keeps temps down and run the 8 "favored" cores at higher clocks than possible if they were all on the same chiplet? Or would the impact on memory performance make this not feasible?

 

[MarkPost]

Windows 11 is a good OS, but my point was that TPU is using a much older version of Windows 11. Windows 11 22H2 launched with a new thread scheduler that works hand in hand with Raptor Lake's new Thread Director.

So TPU using an older version of Windows is obviously not optimal for Raptor Lake compared to using Windows 11 22H2. Even if it makes no difference, it's good review practice to make sure your software is fully updated (including drivers) for benchmarks.

so you are trying to tell us that with 22H2 RPL consumes less power? really?

oh and 21H2 is perfectly updated by Microsoft.

btw, Windows 11, version 22H2 known issues and notifications | Microsoft Learn Lower than expected performance in some games

not solved (apparently) till November 29, 2022—KB5020044 (OS Build 22621.900) Preview - Microsoft Support

 

[Carfax83]

so you are trying to tell us that with 22H2 RPL consumes less power? really?

No, I'm saying that it's generally good practice to have all software including UEFI and drivers updated when doing benchmarking. When I see a review site not using the latest versions of drivers, OS etcetera, it raises a red flag in my mind.

Also, the Windows 11 22H2 has a new scheduler which works in tandem with Intel's thread director 2 in Raptor Lake, which is supposed to more intelligently assign workloads to performance and efficiency cores based on real-time performance metrics; so theoretically, it can affect performance and power efficiency.

btw, Windows 11, version 22H2 known issues and notifications | Microsoft Learn Lower than expected performance in some games

not solved (apparently) till November 29, 2022—KB5020044 (OS Build 22621.900) Preview - Microsoft Support

I clean installed Windows 11 22H2 and I haven't had any issues personally. I know some people definitely had problems though, as is often the case with Windows updates.

 

[Timur Born]

Latest observations on my undervolted 13900K:

- CB23 is stable down to 230W, but in order to reach that low Prime 95 Small FFTs load can become unstable when 16 (HT) unthrottled threads are run only on the 8 P cores (not E cores) at 55x below 220W (and 1 core load likely, too).

- The Arctic LF2 at 100% is not capable of keeping 8 P cores of P95 SFFT AVX2 cool enough at 55x. As a result my Vcore offset needs to increase from -0.84v to -0.82v for stability of this particular load scenario, while single cores (hottest core 5) can hit up to the 100°C temperature limit. CB23 uses up to 245W then.

- For my own setup I am trying the use of TVB now to lower core ratios when cores get too hot. Currently I am evaluating the best balance between lower Vcore for CB23 and lower TVB limit for P95 stability.

- With my combination of AC/DC LLC lowest (Power Saving), CPU LLC 3rd lowest (Medium) and offset of -0.82v (minimum for 8 P core stability) a power limit of around 253W happens to be needed to keep all-core P95 SFFT AVX2 stable. So maybe Intel's 253W limits isn't so arbitrary.

 

[Khanan]

No, I'm saying that it's generally good practice to have all software including UEFI and drivers updated when doing benchmarking. When I see a review site not using the latest versions of drivers, OS etcetera, it raises a red flag in my mind.

You have a point but I don’t think it’s a big deal like someone else already mentioned. He probably didn’t update it so to make it more comparable to the earlier results, this is standard practice.

 

[MarkPost]

No, I'm saying that it's generally good practice to have all software including UEFI and drivers updated when doing benchmarking. When I see a review site not using the latest versions of drivers, OS etcetera, it raises a red flag in my mind.

Also, the Windows 11 22H2 has a new scheduler which works in tandem with Intel's thread director 2 in Raptor Lake, which is supposed to more intelligently assign workloads to performance and efficiency cores based on real-time performance metrics; so theoretically, it can affect performance and power efficiency.

and that supposed feature doesnt matter for power consumption.

No, its not a good practice to have all updated if stability isnt verified. It takes time. So the good practice is to use BIOS, OS, drivers, software probed to be stable and reliable, otherwise you can find incosistences due to bugs we all know are present in new soft. Then you'll cry about that.

btw there are a bunch of reviews showing the same that TPU shows about power consumption in games. The contrary is the outlier.

I clean installed Windows 11 22H2 and I haven't had any issues personally. I know some people definitely had problems though, as is often the case with Windows updates.

Have you compared 21H2 and 22H2 perfomance? I guess not. Its pretty obvious that if a fix was released for 22H2, a bug was present. Its a fact. Here you have an example why its not a good idea to have all soft updated to the latest version.

 

[Timur Born]

Windows 22H2 moves threads of background processes from P cores to E cores if a maximized window of another process is put in focus in the foreground. Only the combination of "Balanced" PLAN and "Best Performance" MODE does not try to do this.

If you are urgently waiting for background processing to finish then use the "Best Performance" power MODE, else use something else to keep P cores free for faster foreground processing and better background efficiency.

Running 7-Zip benchmark shows that E cores at 43x equal P cores at 40x both in performance and power consumption.

 

[moinmoin]

While E cores are comparable to P cores in INT loads, same can't be said for FP loads where E cores fare worse both in performance and in power efficiency as CnC has previously shown. I'd hope Thread Director doesn't forcibly moves such loads from P to E cores.

 

[Timur Born]

While E cores are comparable to P cores in INT loads, same can't be said for FP loads where E cores fare worse both in performance and in power efficiency as CnC has previously shown. I'd hope Thread Director doesn't forcibly moves such loads from P to E cores.

CB23 uses floating-point, as far as I know. A P core at 40x scores 1580 pts. using around 16W, an E core at 43x scores 1180 pts. using about 17.5W.

Windows 11 does not seem to push Cinebench threads to E cores when it runs in the background using the "Balanced" power PLAN + MODE.

 

[Carfax83]

and that supposed feature doesnt matter for power consumption.

In any case, the TPU results don't nullify der8auer's. Both can be correct. The reason why I brought up the Windows version is because the conclusion way well off from der8auer's and I know that Windows 11 22H2 has more intelligent core assignments than previous versions.

Have you compared 21H2 and 22H2 perfomance? I guess not. Its pretty obvious that if a fix was released for 22H2, a bug was present. Its a fact. Here you have an example why its not a good idea to have all soft updated to the latest version.

That bug never affected everyone. I think it mostly affected those that upgraded and not clean installed like I did and was also probably game specific.

 

[Timur Born]

...I know that Windows 11 22H2 has more intelligent core assignments than previous versions.

With "autonomous mode" being enabled P cores with higher maximum frequency are prioritized over those with lower max by Windows 11 22H2. This happens regardless of the Turbo maximum. So a core set to 61x with a Turbo maximum of 58x still gets prioritized over a core with a max of 60x and the same Turbo max of 58x. I (ab)use this to make Windows prioritize cooler and/or more stable cores. Unfortunately this does not seem to work for E cores. I did not check disabling autonomous mode yet.

When Windows/CPU moves threads around to spread heat it prioritizes HT cores (4 ns latency) over physical ones (27-33 ns latency). So a thread first is shifted back and forth between the two logical cores of its physical cores until the core gets too hot and then it is moved to another physical core instead. For P cores this still makes use of above mentioned priorities then.

CPPC on AMD keeps Windows from moving threads around for cooling, which results in temperature hotspots and thus potentially more fan-noise. Curiously Windows offloads *low* background load to the worst core (in terms of VF curve) of the worse CCD, likely because this happens to be one of the most energy efficient cores and low load doesn't interfere too much with the "Frequency Limit Global" (which limits *all* cores at once). I tested this on older Windows versions, though, so things may have changed.

Core Parking - if used - disables HT cores before E cores before P cores on Intel. On my 5900X I mean to remember that it disables the second CCD before HT/SMT cores if CPPC is enabled.

 

[Just Benching]

The review doesn't show any difference in efficiency. At 65w power limit the 7950x consumes 90 while the 13900k consumes 71. If you run the numbers for efficiency, the difference is around 8%.

The 7950x is indeed more efficient in very heavy MT workloads, but the difference is small, 10 to 15% depending on the wattage you are testing. On the other hand in mixed workloads and actual productivvity / content creation, the 13900k is massively faster and more power efficient. Im talking about premiere pro, autocad, photoshop and the likes, not only is the 13900k around 20 to 25% faster, it's also 50 to 80% more efficient.

 

[Hitman928]

The review doesn't show any difference in efficiency. At 65w power limit the 7950x consumes 90 while the 13900k consumes 71. If you run the numbers for efficiency, the difference is around 8%.

The 7950x is indeed more efficient in very heavy MT workloads, but the difference is small, 10 to 15% depending on the wattage you are testing. On the other hand in mixed workloads and actual productivvity / content creation, the 13900k is massively faster and more power efficient. Im talking about premiere pro, autocad, photoshop and the likes, not only is the 13900k around 20 to 25% faster, it's also 50 to 80% more efficient.

This video was posted before, but maybe in another thread. I still think there is a problem with his setup on the AMD side, I haven't seen any other reviewer report such high idle CPU power for Zen4. It will be higher than other monolithic chips for sure, but it shouldn't be that much higher. It also doesn't really agree with pretty much every other reviewer I've seen that tests for efficiency, including Anandtech. True, Anandtech didn't test the same way/software for the most part, but if you look at Anandtech's 35W and 65W results, they would make absolutely no sense whatsoever (there would be no way Zen4 would be beating RPL, even with the TDP vs PPT issue on the AMD side) if the Zen4 CPUs were truly using so much power just at idle as that video suggests. There is clearly something wrong with his setup on the AMD side or the way he is measuring.

 

[Just Benching]

This video was posted before, but maybe in another thread. I still think there is a problem with his setup on the AMD side, I haven't seen any other reviewer report such high idle CPU power for Zen4. It will be higher than other monolithic chips for sure, but it shouldn't be that much higher. It also doesn't really agree with pretty much every other reviewer I've seen that tests for efficiency, including Anandtech. True, Anandtech didn't test the same way/software for the most part, but if you look at Anandtech's 35W and 65W results, they would make absolutely no sense whatsoever (there would be no way Zen4 would be beating RPL, even with the TDP vs PPT issue on the AMD side) if the Zen4 CPUs were truly using so much power just at idle as that video suggests. There is clearly something wrong with his setup on the AMD side or the way he is measuring.

I don't know about idle power but intel beating amd in mixed workloads isn't anything new, the same was true between ALD and zen 3.

Anandtech was testing heavy MT, and even in those scenarios the difference in efficiency was tiny (8%), the zen 4 part was consuming considerably more power than the intel counterpart when both were limited at the same wattage.

For example

At 65w the difference in efficiency might look huge, but then you realize this

The amd part was drawing way more power. If you actually do the math 22911 / 71 = 322pts vs 346 pts for the amd part. So the AMD part is 7.4% more efficient at 65w in heavy MT

 

[Hitman928]

I don't know about idle power but intel beating amd in mixed workloads isn't anything new, the same was true between ALD and zen 3.

Anandtech was testing heavy MT, and even in those scenarios the difference in efficiency was tiny (8%), the zen 4 part was consuming considerably more power than the intel counterpart when both were limited at the same wattage.

For example

At 65w the difference in efficiency might look huge, but then you realize this

The amd part was drawing way more power. If you actually do the math 22911 / 71 = 322pts vs 346 pts for the amd part. So the AMD part is 7.4% more efficient at 65w in heavy MT

The main criticism I and others had for Anandtech's article is that they didn't take actual power measurements for the applications they benchmarked except yCruncher. Problem is, both Intel and AMD allow their CPUs to draw above TDP and how much can depend on motherboard settings and application. It is probably safe to assume that both are drawing the same in Cinebench as yCruncher for lower power levels, but this might not be true at the high end. In general, I agree with your point that RPL is closer in efficiency to Zen4 on the desktop than they get credit for. Part of the problem for Intel, though, is that at low power levels (per core power) is where Zen4 shines which is where the trouble will be for Intel because that is what will allow them to dominate in performance and efficiency in the laptop and server markets, where their customers care much more about this.

As far as the idle power goes, I mentioned it because in the video you posted, his minimum power recorded for the Zen4 CPUs are inexplicably high which , when testing mixed work loads, will hurt Zen 4's efficiency performance.

 

[Thunder 57]

The review doesn't show any difference in efficiency. At 65w power limit the 7950x consumes 90 while the 13900k consumes 71. If you run the numbers for efficiency, the difference is around 8%.

The 7950x is indeed more efficient in very heavy MT workloads, but the difference is small, 10 to 15% depending on the wattage you are testing. On the other hand in mixed workloads and actual productivvity / content creation, the 13900k is massively faster and more power efficient. Im talking about premiere pro, autocad, photoshop and the likes, not only is the 13900k around 20 to 25% faster, it's also 50 to 80% more efficient.

I'll keep it short since I don't care and you are stirring up nonsense for the sake of being a contrarian.

Well, hell, there really isn't anything to say . You might as be that HS guy. A total troll. Guess I'll get in trouble for this but so be it.

 

[Just Benching]

I'll keep it short since I don't care and you are stirring up nonsense for the sake of being a contrarian.

Well, hell, there really isn't anything to say . You might as be that HS guy. A total troll. Guess I'll get in trouble for this but so be it.

Do you disagree with the data i presented or you just dont like them? Cause resorting to name calling aint going to change the data

 

[Markfw]

Do you disagree with the data i presented or you just dont like them? Cause resorting to name calling aint going to change the data

Anybody can present crap data, skewed data or cherry picked data. We have presented you with many examples proving you are wrong and trolling, and you still think you are right.

 

[Just Benching]

Anybody can present crap data, skewed data or cherry picked data. We have presented you with many examples proving you are wrong and trolling, and you still think you are right.

I used the anandtechs review numbers. Are they crap skewed and cherry picked?

I dont know who are the "we" you are talking about, but im certain you haven't presented me anything cause I dont know you and ive never talked to you again in my life. So you are clearly wrong.

 

[DrMrLordX]

The review doesn't show any difference in efficiency. At 65w power limit the 7950x consumes 90 while the 13900k consumes 71. If you run the numbers for efficiency, the difference is around 8%.

That video is kinda stupid, and you obviously don't understand AMD's power limits vs. Intel's.

 

[Markfw]

I used the anandtechs review numbers. Are they crap skewed and cherry picked?

I dont know who are the "we" you are talking about, but im certain you haven't presented me anything cause I dont know you and ive never talked to you again in my life. So you are clearly wrong.

I say "we" as anandtech forum members, and not just one person has said you are wrong.

 

[Just Benching]

That video is kinda stupid, and you obviously don't understand AMD's power limits vs. Intel's.

What video, my numbers are taken from anandtechs review this thread is about. At 65w power limit the zen 4 part drew considerably more power than the rpl, making the difference in efficiency just 7.5%. Decent, but not huge, nothing to write home about

About the video, what part is stupid?

 

[DrMrLordX]

What video, my numbers are taken from anandtechs review this thread is about.

The one you linked?

At 65w power limit the zen 4 part drew considerably more power than the rpl, making the difference in efficiency just 7.5%. Decent, but not huge, nothing to write home about

Again you don't understand the difference between AMD and Intel power limits. When you set a TDP of 65W for modern AMD CPUs, it allows a maximum of 1.35xTDP in power draw if sufficient cooling is available. With Intel is is far more complicated, involving PL1-PL4 numbers and tau values. You can (almost) hard limit Intel power draw by setting PL1 = PL2 = 65W, while with AMD if you want to limit it to 65W, you would either set TDP to ~48W or limit PPT (via EDC and TDC if necessary).

So naturally the AMD system will draw more power if you just set TDP to 65W and let her rip while hard limiting PL values on the Intel system. In the real world, where people have measured perf/watt of Raphael extensively, Raphael obliterates Raptor and Alder Lake in the efficiency department when both CPUs are actually drawing the same amount of power.

 

[Just Benching]

I say "we" as anandtech forum members, and not just one person has said you are wrong.

Gotcha, you just dont like the numbers