AMD vs Intel - Power efficiency

[Redaktion]

The following chart shows the achieved points per watt in the Cinebench R15 multicore benchmark test (1 run). The energy consumption of the entire laptop system is measured with the external monitor connected (int. panel disabled).

https://www.notebookcheck.com/AMD-vs-Intel-Power-efficiency.581245.0.html

[eloyard]

So judging by the difference between your benchmarks and what Intel claimed in their slide, if alder lake manages to reach
Power/performance parity with mobile zen 3 that could be considered success.

/Golfclap GG Intel

[Russel]

So judging by the difference between your benchmarks and what Intel claimed in their slide, if alder lake manages to reach
Power/performance parity with mobile zen 3 that could be considered success.

/Golfclap GG Intel

Not exactly. Alderlake should compete with zen 3+ rather than Zen 3.
But since Zen 3+ wouldn't really offer all that much performance gains over zen 3 (amd's claims are yet to be verified, and they used a 28W configuration of the U series), Alderlake could really be considered a winner. Buuuuuuuut intel has multiple times the resources, they were lazy for over 10 years and milked the customers with quad-cores and even dual-cores. Soooooo.   They don't really deserve any head pats.

[thevisi0nary]

Jesus Christ, that per device chart for cinebench power consuption is one of the worst designed charts I've ever seen. How are you supposed to read this? Match one of the 4 shades of pink lines to one of the 4 shades of pink fonts for the devices? They're not even in order!

[ArsLoginName]

As a journalist, please dig a little deeper regarding Intel and Alder Lake mobile. Go through their footnotes from CES 2022 presentation and find the following:

1 - The AMD system had 16 GB RAM while both Intel systems had 32 GB (see footnotes)

2 - According to the link Intel provides in the footnotes, [Ref 1], these systems were tested on Win11 Build 22000.222 which was out sometime between Oct 4th (Win 11 launched Oct 5th, Build 22000.194) and Oct 12th (Build 22000.258). [Ref 2] Hence the testing most likely contained the AMD cache bug so the actual 5900HX performance is probably better than shown because the cache bug fix was in Build 22000.282 on Oct 21/22 and the interim fix worsened the cache and performance issue.

3 - If you look at the 2 slides for Content Creation lead for the new 12th Gen i9 H series processor, it is between 9-44% better relative to the 11th Gen i9 Tiger Lake part. The benchmarks for both of the i9's are run in a MSI Raider 17" laptop which has been shown to consume desktop levels of power. From your own review of the i9-11980HK in the Raider [Ref 3] : "MSI configures the latter with a short-term value (PL2) of 200 watts and a long-term value (PL1) of 91.75 watts". So great. Intel's 12th gen beats their 11th gen by around 10% in 3 tests (Lightroom, AutoCad, and Revit) and 30-44% in Premiere and Blender, which strongly depend on the number of threads. But note, the "Intel 7" process alone should provide a 10-15% performance per watt increase according to Intel's own guidance [Ref 4]. Intel lists max turbo (PL2 ?) power as 115 W . Is this up from the 95 W of the 11th Gen?. So are these really much better than the 11th gen at same power?

4 - The performance-power scaling graph Intel shows is completely out of whack of what is observed by their own 11th Gen Tiger Lake Reference platform UNLESS the i9-11980HK is substantially faster than the i7-11800H. In Techspot's i7-11800H review [Ref 5], the i7-11800H is roughly 20-25% slower than the 5900HX from 35-65 W while Intel is now showing the i9-11980HK being from 0-10% faster than the 5900HX based upon the slope of the graphs. Is the i9-11980HK (2.6 GHz base freq, 5.0 GHz turbo) really 20%+ faster than the i7-11800H (2.3 GHz base, 4.6 GHz turbo)? Intel's own frequency numbers suggest otherwise at the base 45 W TDP.

(1) intel.com/performanceindex
(2) docs.microsoft.com/en-us/window ... nformation
(3) www.notebookcheck.net/MSI-GE76- ... 970.0.html
(4) www.anandtech.com/show/16823/in ... ib-foveros
(5) ww.techspot.com/review/2262-in ... i7-11800h/

[Kdo]

This is misleading.

Presenting results in performance per watt suggests a linear correlation between power and performance. As if 100 perf for 100 W is equivalent to 50 perf for 50 W.

We all know this is completely untrue. To begin with, a system which reaches 50 perf at 50 W may not be able to consume 100W, let alone produce 100 perf, so it's not "the same". More importantly, the performance per watt curve is asymptotic, not linear. In the first few watts it may be close to linear trend, but then it quickly becomes "let's consume a lot more power for a tiny bit more performance".

Showing AMD with a hugely better perf per watt suggests that, at same wattage, the AMD is hugely more powerful. But this is untrue, because both systems where NOT measured at same wattage. That's the misleading part.

What we know of AMD Zen3 is that it offers a little bit less performance than Tiger Lake for a lot less energy consumption, hence the better "efficiency". But that's mostly because they occupy different positions in the curve. When setting Tiger Lake consumption to match the Zen3 one, the differences become much smaller. At ideal wattage for Zen3, Zen3 tends to be a little more powerful than Tiger Lake, not vastly more. And at high wattage, well, Zen3 is unable to take advantage of high wattage, so Tiger Lake becomes better while Zen3 performance curve is essentially flat.

This suggest that a single "number" is a misleading representation of reality, and that a more thorough evaluation should present a curve, with different watt / perf trade off.

Now I understand that it's a lot more work, so it's easier to produce a single number. In which case, at least ensure that all systems use the same wattage (except of course those that are unable to consume that much). This would be a more "fair" representation of efficiency, even though the selection of the reference wattage is itself a criteria by which it's easy to make one system or another the "winner" of the efficiency race.

[Vaidyanathan]

@Kdo,

Fair point, but like you yourselves said that process can inherently make one device more efficient than the other, so we have to make the best of available data across existing benchmarks. Point noted, though. I have added a Cinebench R15 performance chart so that the actual performance numbers are also available.

[Vaidyanathan]

@ArsLoginName

Thank you for those pointers and the useful references. Much appreciated. I didn't particularly talk much about Alder Lake as it's hard to rely on marketing numbers when you are discussing something based on hard data. Eventually, we will get enough data for ADL laptops through the year, so we will get a better idea as to how things are working out for Intel. Also, by then hopefully we'll actually have reliable Windows 11 performance on AMD systems.

[Vaidyanathan]

Jesus Christ, that per device chart for cinebench power consuption is one of the worst designed charts I've ever seen. How are you supposed to read this? Match one of the 4 shades of pink lines to one of the 4 shades of pink fonts for the devices? They're not even in order!

Lol. We have changed some colors and the chart opacities for better legibility. Hope this is better now. Hover over each individual device for better isolating its graph. :D

[MOFO]

Not too many people really care for power efficiency only what performance a CPU brings to the table.

[Spunjji]

"Not too many people really care for power efficiency only what performance a CPU brings to the table"
Power efficiency is the most important metric in a notebook CPU. There's not much point having the best peak performance if most devices can't sustain it, the ones that can need to be thick and heavy and/or have obscenely loud fans, and/or that performance disappears when you're running on battery power.

[Spunjji]

I just wanted to add my gratitude to the Notebookcheck team for doing this testing and presenting the results. I had already received this impression (that Intel's slender performance lead with their 11th gen comes at the expense of power efficiency, and is only sustainable in designs with powerful cooling) by reading between the lines of various reviews, but it's very helpful to see it laid out here.

Personally, I am not very optimistic of improvements with Alder Lake. I suspect the situation in Cinebench will improve substantially, as those E cores will provide a "wider" design allowing for similar performance at much lower clock speeds (and thus lower voltages), but I'm not sure that will translate as well into tasks that aren't embarrassingly parallel. The E cores don't seem to be much more power-efficient than the P cores running at a lower clock speed.

[Vaidyanathan]

@Spunjji,
Thank you for the encouragement and feedback and glad that you've liked the article.
Let's see what Intel has in store for Alder Lake on mobile. If any, the U-series should be quite interesting too.

[Alder Zen]

Even after 2 years Intel still can't compete with TSMC's 7N process and Intels "Intel 7" process is worse.

[Flaschenhans]

Many thanks for the analyses! You had collected a lot of data. It is worthwhile to look at them again as a whole. :-)

I want to elaborate on @Kdo's comment, because the analysis has not considered (nor mentioned) another aspect: power scaling, i.e., an increased power consumption by X% does not lead to the same performance increase of X%, see techspot.com/review/2262-intel-core-i7-11800h and youtube.com/.... The lower the power consumption, the better the energy efficiency (typical for a processor).

Why is this relevant to your analysis? Although they are all "45W" CPUs, the Intel CPUs generally run at much higher wattages (see your 'Power Consumption Cinebench R15' interactive graph). So your analysis has a bias for AMD, and against Intel. Accordingly, a fairer analysis would be to compare CPUs for specific wattages (or wattage ranges); see links above.

I did this over the last year using your data for the i7-11800H and R7-5800H, since you generally report the PL1 (Sustained) wattage numbers (but not always correctly; sometimes I had to readjust them using the screenshots in the individual tests if the power consumption is apparent, otherwise I ignored the test as a data point). Accordingly, I arrive at these values (as of Nov. 2021, average values based on the averages from your Cinebench R15 Multi endurance tests, i.e. not the maxima of the often first run):

Watt-range (+/-5)  i7-11800H  R7-5800H  Energy efficiency (R7 vs. i7)351405193838%451675208925%551790200412%651855no data-75195221239%85206220520%

What does that mean? The R7-5800H is only over 30% more energy efficient than the i7 as long as both run at less than 45W. With increasing wattage, the R7's advantage gradually dwindles. This is also confirmed in the analyses shown in the links above (i.e., the power scaling of the R7 is worse than that of the i7, at least above 50W).