Intel Core i9-12900HK performance and power efficiency comparison: Significant gains with potential to put even upcoming AMD offerings at a disadvantage

[Redaktion]

Our first Alder Lake-H performance numbers are in, and enthusiasts seem to be in for a treat. We pit the Core i9-12900HK in the MSI GE76 Raider against Tiger Lake-H, AMD Cezanne Zen 3, and Apple M1 Max SoCs to see what kind of improvements one can expect with Intel's latest mobile offering. From what we see so far, Intel has finally been able to thwart both AMD and Apple in their tracks but at some expense of power.

https://www.notebookcheck.net/Intel-Core-i9-12900HK-performance-and-power-efficiency-comparison-Significant-gains-with-potential-to-put-even-upcoming-AMD-offerings-at-a-disadvantage.595060.0.html

[Right.... Bs]

Already calling BS much like I had about the rumor Samsung was dropping AMD.

[A]

Power efficiency needs to be tested at the same wattage. Alder lake will be the most efficient chip out there at 45-60W power ranges.

[Vaidyanathan]

Power efficiency needs to be tested at the same wattage. Alder lake will be the most efficient chip out there at 45-60W power ranges.

Fair point. We are contemplating on this, but one thing needs to be pointed out. These are data from device reviews wherein the idea is not to just restrict the assessment to the CPU alone but to the device as a whole. Also, there is the possibility of an inherent bias for selecting a certain wattage as one value could automatically put a system at an advantage or a disadvantage depending on how it can handle the load.

[Rob Stan]

6980HX is the top of the line Rembrandt SKU, not 6900HX, I could get why last gen everyone was testing the 5900HX as 5980HX was almost nonexistent in laptop designs, but this year that does not seem to be the case.

That being said, all of these smorgasbord results in the charts are extremely confusing, if not downright misleading given they don't give any info on power setting (and no info on power measurements). It's like everyone reviewing Intel stuff has decided to give them a pass on their PL1=PL2 (and various other) scumminess. That might fly on desktop (if you at leat warn the potential DIY buyers to get a beefy AIO) but it under no circumstance be ignored on laptop.

[Rob Stan]

Power efficiency needs to be tested at the same wattage. Alder lake will be the most efficient chip out there at 45-60W power ranges.

Really? The exact opposite was true with Tiger Lake-H. I don't see how it will be vastly different this time, especially with the PL1=PL2=115W and temporary TDPup to 130W ADL mobile has.

Unlike Cezanne, both TGL and ADL scale up performance nearly linearly with more voltage and watts, so performance will also drop drastically at lower TDP which means they barely have a sweet spot of performance to power to begin with.

Considering the whole big-little approach, the fact that this behaviour seems to repeat itself over from the previous gens makes the whole thing look sad.

Apple (node advantage is a big help in this), and to a slightly lesser degree, AMD, continue to be the much more efficient alternatives.

[Vaidyanathan]

That being said, all of these smorgasbord results in the charts are extremely confusing, if not downright misleading given they don't give any info on power setting (and no info on power measurements).

Take a look at the linked review for detailed power measurements.

[_MT_]

Unlike Cezanne, both TGL and ADL scale up performance nearly linearly with more voltage and watts, so performance will also drop drastically at lower TDP which means they barely have a sweet spot of performance to power to begin with.

Clearly, that's not true. Not only it makes no sense from physics standpoint, data in the article clearly shows otherwise. Just contrast HK with K with almost half the efficiency. It's pretty much guaranteed that efficiency will drop as frequency goes up beyond 3 GHz. I doubt they have managed to push the sweet spot that high. Traditionally, it would be more like 2 GHz. The further you go, the worse it gets.

I would like to see power- or performance-normalized numbers. If I need efficiency, if I need battery life, I can sacrifice performance. So, this is relevant information, at least to me.

[_MT_]

The relatively low frequency is part of the reason why is M1 so efficient.

[well]

The performance per watt measurement is not done correctly.
Why show such metric in cinebench r15 where it's not running natively on Apple Silicon?

The measurement should be in cinebench r23 with all chips running natively.

[ArsLoginName]

You dropped the ball on this one. You can't proclaim 'efficiency crown' for Alder Lake and such when you don't benchmark for it. Anandtech did and found "this top-of-the-line" completely non-thermally restrained desktop replacement processor performed at 3494 points in Cinebench R20 at 30 W. Your own database shows an *average* score of 3303 for a Ryzen 5800U. That's a 5.8% difference between these processors with 3 of the Ryzen 5800u scores in 13" notebook chassis. I read here daily for the honest truth and not sensationalism  of other sites. Stay objective.

[Vaidyanathan]

You dropped the ball on this one. You can't proclaim 'efficiency crown' for Alder Lake and such when you don't benchmark for it. Anandtech did and found "this top-of-the-line" completely non-thermally restrained desktop replacement processor performed at 3494 points in Cinebench R20 at 30 W. Your own database shows an *average* score of 3303 for a Ryzen 5800U. That's a 5.8% difference between these processors with 3 of the Ryzen 5800u scores in 13" notebook chassis. I read here daily for the honest truth and not sensationalism  of other sites. Stay objective.

Didn't get you. Where have I proclaimed any efficiency crown? I only stated it's good in performance and needs more work in terms of efficiency.

[Vaidyanathan]

The performance per watt measurement is not done correctly.
Why show such metric in cinebench r15 where it's not running natively on Apple Silicon?

The measurement should be in cinebench r23 with all chips running natively.

Historically, we have used power consumption from the wall data while doing a Cinebench R15 run on an external monitor. So, it is not possible to change for just one SoC right away. I agree this is a limitation of the test. Having said that a CB15 multi render is not a huge deal for the M1 Max with 10 CPU cores even if you account for the Rosetta overhead. You can see that by comparing the deltas between the M1 Max and the 12900HK between CB15 Multi and CB23 Multi runs. They are not way off.

[Jakemaster]

Excellent summary analysis in this article. The best of the three I've read today.  Keep up the good work!

[Vaidyanathan]

Excellent summary analysis in this article. The best of the three I've read today.  Keep up the good work!

Thank you Jakemaster! Glad you found it useful :)