Info FYI: Comparison of Alder Lake's performance vs efficiency cores (Geekbench 5)

[horshack]

I was curious about exactly how the Intel 12th gen efficiency cores compare to the performance cores, so I wrote a script that runs Geekbench 5 against each individual core. I chose Geekbench for this test because it offers a nice mix of real-world algorithms rather than an arbitrary synthetic mix of computations. Here are the results for an i7-12700h running in an Asus Vivobook S 14X.

Summary of results
-----------------------------------------------
Core # 0: Average=1771.00  *** Baseline ***
Core # 1: Average=1770.50  vs Baseline:  99.97%
Core # 2: Average=1767.33  vs Baseline:  99.79%
Core # 3: Average=1749.00  vs Baseline:  98.75%
Core # 4: Average=1844.16  vs Baseline: 104.13%
Core # 5: Average=1843.66  vs Baseline: 104.10%
Core #12: Average=1091.66  vs Baseline:  61.64%
Core #13: Average=1097.66  vs Baseline:  61.97%
Core #14: Average=1096.50  vs Baseline:  61.91%
Core #15: Average=1098.50  vs Baseline:  62.02%
Core #16: Average=1097.83  vs Baseline:  61.98%
Core #17: Average=1095.50  vs Baseline:  61.85%
Core #18: Average=1097.16  vs Baseline:  61.95%
Core #19: Average=1096.33  vs Baseline:  61.90%

Cores #0-5 are the six native performance cores (hyperthreaded cores excluded); cores #12-19 are the eight efficiency cores. Core #0 is the 100% baseline.

Based on these results the efficiency cores are 62% as fast as the performance cores on a i7-12700h.

Note how cores #4 and #5 are faster than the other four performance cores - these are the "favored cores" on this particular CPU die. See Intel Turbo Boost Max Technology 3 for details.

Here is a link to my GitHub repository if anyone would like to run this test on their setup as well. It runs under Linux.

https://github.com/horshack-dpreview/GeekbenchAutomationTools

 

[eek2121]

FYI Anandtech has a controlled test of the E cores as well. Hit up the bench section and look at their results. Interesting stuff to be sure, basically, Skylake with improved FP/AVX (aka "Gracemont") does very well in modern benchmarks.

 

[Markfw]

FYI Anandtech has a controlled test of the E cores as well. Hit up the bench section and look at their results. Interesting stuff to be sure, basically, Skylake with improved FP/AVX (aka "Gracemont") does very well in modern benchmarks.

I looked on bench, and I se CPUs, but not just the E-cores. Can you link the actual e-cores bench ?

 

[horshack]

FYI Anandtech has a controlled test of the E cores as well. Hit up the bench section and look at their results. Interesting stuff to be sure, basically, Skylake with improved FP/AVX (aka "Gracemont") does very well in modern benchmarks.

Thanks, missed that. They isolated cores using affinity masks, whereas I'm removing the cores' accessibility from the kernel. Probably doesn't make a difference in practice but I'll look into it.

I looked on bench, and I se CPUs, but not just the E-cores. Can you link the actual e-cores bench ?

I found this after @eek2121 mentioned it:

https://www.anandtech.com/show/1704...ybrid-performance-brings-hybrid-complexity/10

 

[nicalandia]

Impressive. Thanks for sharing

 

[igor_kavinski]

Summary of results
-----------------------------------------------
Core # 0: Average=1771.00  *** Baseline ***
Core # 1: Average=1770.50  vs Baseline:  99.97%
Core # 2: Average=1767.33  vs Baseline:  99.79%
Core # 3: Average=1749.00  vs Baseline:  98.75%
Core # 4: Average=1844.16  vs Baseline: 104.13%
Core # 5: Average=1843.66  vs Baseline: 104.10%
Core #12: Average=1091.66  vs Baseline:  61.64%
Core #13: Average=1097.66  vs Baseline:  61.97%
Core #14: Average=1096.50  vs Baseline:  61.91%
Core #15: Average=1098.50  vs Baseline:  62.02%
Core #16: Average=1097.83  vs Baseline:  61.98%
Core #17: Average=1095.50  vs Baseline:  61.85%
Core #18: Average=1097.16  vs Baseline:  61.95%
Core #19: Average=1096.33  vs Baseline:  61.90%

Pretty nifty script! The E-core score tells me that it's comparable to my Broadwell with 128MB eDRAM. If these e-cores were on their own die, I guess I could learn to like them.

 

[moinmoin]

@horshack would it be possible to extend the script with a RAPL energy usage reading in joules? I'm using perf stat -a -e "power/energy-pkg/" but there likely are better ways to get the numbers from RAPL.

 

[Hulk]

Very nice work. Thanks. It'd be great to know clocks during the run.

 

[Roland00Address]

Do we know what mhz the efficiency and the performance cores are running at?

Wiki says the max turbo is 3500 mhz for e and 4700 mhz for p cores. But I do not know what the speed your cores are running since it is a laptop and previous generations max turbo was for 1 core while multiple cores can hit lower mhz.

3500 vs 4700 mhz is 74.5% just due to clock speed.

 

[zir_blazer]

You may be interesed in my own benchmarks on Cinebench R23.

 

[DrMrLordX]

If these e-cores were on their own die, I guess I could learn to like them.

Just noticed your comment. Check out the elusive Alder Lake-N:

Intel's Alder Lake-N details leaked (low-cost, low-power chips that will succeed Jasper Lake) - Liliputing

Intel's Alder Lake-N details leaked (low-cost, low-power chips that will succeed Jasper Lake)

liliputing.com

Looks like they're going to Chromebooks:

Alder Lake-N Chromebooks may seriously leapfrog performance for affordable devices

For years at this point, Chromebooks with Intel silicon inside have fallen on one of two sides of the coin: large core or small core. That's a simplistic way to put it, but in general, this is how we've seen Intel approach Chromebooks over the past 6 or 7 years. As an example, as 4th-gen...

chromeunboxed.com

 

[eek2121]

Just noticed your comment. Check out the elusive Alder Lake-N:

Intel's Alder Lake-N details leaked (low-cost, low-power chips that will succeed Jasper Lake) - Liliputing

Intel's Alder Lake-N details leaked (low-cost, low-power chips that will succeed Jasper Lake)

liliputing.com

Looks like they're going to Chromebooks:

Alder Lake-N Chromebooks may seriously leapfrog performance for affordable devices

For years at this point, Chromebooks with Intel silicon inside have fallen on one of two sides of the coin: large core or small core. That's a simplistic way to put it, but in general, this is how we've seen Intel approach Chromebooks over the past 6 or 7 years. As an example, as 4th-gen...

chromeunboxed.com

Nice. I am curious how it will fair against Zen 2 parts.

IMO Gracemont (or Atom in general) doesn’t get enough love at Intel.

 

[DrMrLordX]

IMO Gracemont (or Atom in general) doesn’t get enough love at Intel.

Obviously they take it pretty seriously if they're using it in their top-end desktop parts. But it seems like an awkward fit for now. A Gracemont-only product would be a nice successor to Tremont and a major step forward for Atom in general.

 

[igor_kavinski]

A Gracemont-only product would be a nice successor to Tremont and a major step forward for Atom in general.

But they would clock it pretty low to make it suitable for fanless operation. Maybe something like a GB5 ST score of 600-700. But I guess acceptable even if that low. My Ivy Bridge Thinkpad scores are similar.

 

[DrMrLordX]

But they would clock it pretty low to make it suitable for fanless operation. Maybe something like a GB5 ST score of 600-700. But I guess acceptable even if that low. My Ivy Bridge Thinkpad scores are similar.

Gracemont should be able to stay within Tremont's power envelope at a peak clockspeed of ~3.2 GHz for top-bin parts. Obviously that's going to change based on the power targets, but you get the idea.

 

[igor_kavinski]

I'm just wondering what they are waiting for. Maybe planning to mass produce ADL-N on Intel 4 for lower power consumption and more profit per wafer? They've had GM cores since 2021. Why sit on something that could be a worthy competitor to cheap AMD chromebooks/netbooks?

 

[Hulk]

I thought Gracemont was for area efficiency, not energy?
In addition, Big.Little in ADL is foreground/background apps, right?


Intel’s 12th-gen Core processors, code-named Alder Lake, incorporate two different types of CPU cores: high performance CPU cores based on the new Golden Cove Core architecture and energy efficient chips based on the new Atom-based Gracemont architecture. The idea is that most Alder Lake chips can use the efficient cores to save power when they’re all that’s needed, use the performance cores when you need the extra power, and use all cores together for an even bigger boost in multi-core performance.

 

[DrMrLordX]

I thought Gracemont was for area efficiency, not energy?

It is in Alder Lake-S, since they push the clocks too high (basically). It's also because Golden Cove is actually pretty energy efficient as well. It just costs too much in silicon area.

As a successor to Tremont, Gracemont could be very efficient assuming Alder Lake-N (and other Gracemont-based products) keep clocks low enough.

 

[igor_kavinski]

I would like to know what happens in pure MT workloads when P-cores finish threads faster than the E-cores? Are the pending threads paused on the E-cores and moved to the freed-up P-cores or do they continue running on the E-cores till completion? If it's the latter, I suppose that's leaving considerable computing power on the table.

 

[Zucker2k]

I would like to know what happens in pure MT workloads when P-cores finish threads faster than the E-cores? Are the pending threads paused on the E-cores and moved to the freed-up P-cores or do they continue running on the E-cores till completion? If it's the latter, I suppose that's leaving considerable computing power on the table.

This is probably why Intel is clocking the e-cores even higher this time around. If Gracemont's efficiency on Alderlake-S wasn't mind-blowing because it shared the same power plane with Golden Cove, things should definitely look better now in that category. On ADL-S 4 Gracemont cores consumes about the same power on full load (3.9Ghz) as 1 Golden Cove core does in multithreading - 25w.

 

[LightningZ71]

I'm just wondering what they are waiting for. Maybe planning to mass produce ADL-N on Intel 4 for lower power consumption and more profit per wafer? They've had GM cores since 2021. Why sit on something that could be a worthy competitor to cheap AMD chromebooks/netbooks?

In a word, Money. Intel still has limited throughput on their "Intel 7" node. They are prioritizing higher ASP products that will net them the most profit per wafer at present. We know that AlderLake N is a thing, it just hasn't hit any sort of appreciable volume. It likely won't until Intel starts moving product on Intel 4, or they bring online more volume on Intel 7. Keep in mind, Alderlake N and the rest of the Celeron and Pentium line are not only low-power, but also bargain basement volume products, filling the sockets of sub $300 laptops and chromebooks and tiny PCs attached to monitors and hidden away on corporate drone desks all over the world

 

[Hulk]

I was curious about exactly how the Intel 12th gen efficiency cores compare to the performance cores, so I wrote a script that runs Geekbench 5 against each individual core. I chose Geekbench for this test because it offers a nice mix of real-world algorithms rather than an arbitrary synthetic mix of computations. Here are the results for an i7-12700h running in an Asus Vivobook S 14X.

Summary of results
-----------------------------------------------
Core # 0: Average=1771.00  *** Baseline ***
Core # 1: Average=1770.50  vs Baseline:  99.97%
Core # 2: Average=1767.33  vs Baseline:  99.79%
Core # 3: Average=1749.00  vs Baseline:  98.75%
Core # 4: Average=1844.16  vs Baseline: 104.13%
Core # 5: Average=1843.66  vs Baseline: 104.10%
Core #12: Average=1091.66  vs Baseline:  61.64%
Core #13: Average=1097.66  vs Baseline:  61.97%
Core #14: Average=1096.50  vs Baseline:  61.91%
Core #15: Average=1098.50  vs Baseline:  62.02%
Core #16: Average=1097.83  vs Baseline:  61.98%
Core #17: Average=1095.50  vs Baseline:  61.85%
Core #18: Average=1097.16  vs Baseline:  61.95%
Core #19: Average=1096.33  vs Baseline:  61.90%

Cores #0-5 are the six native performance cores (hyperthreaded cores excluded); cores #12-19 are the eight efficiency cores. Core #0 is the 100% baseline.

Based on these results the efficiency cores are 62% as fast as the performance cores on a i7-12700h.

Note how cores #4 and #5 are faster than the other four performance cores - these are the "favored cores" on this particular CPU die. See Intel Turbo Boost Max Technology 3 for details.

Here is a link to my GitHub repository if anyone would like to run this test on their setup as well. It runs under Linux.

https://github.com/horshack-dpreview/GeekbenchAutomationTools

I made a similar analysis using more caveman methods.

For my 12700K I took the CB R23 ST P score and turned it into points/GHz, which came to be 389.
Next I subtracted the 8+0 score from the 8+4 score to isolate the E's contribution. I arrived at a points/GHz of 242 for the E core.
According to this calculation, which is frequency independent in CB R23 one E core is worth 61% of a P (without hyperthreading).
Right on your number if your calcs take frequency into account. If you didn't make it iso-frequency then it could be a different in CB vs Geekbench?

 

[OneEng2]

Wow. Lots of super interesting information.

My question (which I am sure one of you super knowledgeable guys will know), is how much die space does a P core take compared to an E Core? Based on the @zir_blazer link, a single P core scores 2.7 times that of a single E core.

From here: https://locuza.substack.com/p/die-walkthrough-alder-lake-sp-and

It looks like a P core cost you about 10.5mm^2 (on average including the L3 sharing). The article doesn't show the gracemont core complex including L3 number; however, it does say that the core complex is 15% larger than a single Golden Cove core. I am willing to say that the L3 would also be about that much more compared to that shown on each pair of Golden Cove cores. Extrapolating that out would give a per core die size for Gracemont cores (sharing the cost of the L3 as I did with the Golden Cove cores). I got that figured at only 2.29mm^2 per Gracemont core (pretty tiny little things).

So, a Golden Cove core costs you 4.6 times the die space and performs 2.7 times as good..... in this benchmark at least.

So I would conclude that Intel has made a good bet on the "big-little" architecture from a die size perspective. Sadly, they are still at a big yield disadvantage to AMD's architecture because their single die is so big compared to AMD's much smaller CPU tiles.

It will be interesting to see what shakes out when Intel has tiles and AMD has big little .

Great thread.

 

[Hulk]

Wow. Lots of super interesting information.

My question (which I am sure one of you super knowledgeable guys will know), is how much die space does a P core take compared to an E Core? Based on the @zir_blazer link, a single P core scores 2.7 times that of a single E core.

From here: https://locuza.substack.com/p/die-walkthrough-alder-lake-sp-and

It looks like a P core cost you about 10.5mm^2 (on average including the L3 sharing). The article doesn't show the gracemont core complex including L3 number; however, it does say that the core complex is 15% larger than a single Golden Cove core. I am willing to say that the L3 would also be about that much more compared to that shown on each pair of Golden Cove cores. Extrapolating that out would give a per core die size for Gracemont cores (sharing the cost of the L3 as I did with the Golden Cove cores). I got that figured at only 2.29mm^2 per Gracemont core (pretty tiny little things).

So, a Golden Cove core costs you 4.6 times the die space and performs 2.7 times as good..... in this benchmark at least.

So I would conclude that Intel has made a good bet on the "big-little" architecture from a die size perspective. Sadly, they are still at a big yield disadvantage to AMD's architecture because their single die is so big compared to AMD's much smaller CPU tiles.

It will be interesting to see what shakes out when Intel has tiles and AMD has big little .

Great thread.

A 4 E core "cluster" is a little larger than a P core. Yes, E's are there for area efficiency, not power efficiency!
And you are also right in that 1 P is equal to just under 3E's from a compute point of view. They are there for maximum compute for highly parallel tasks where they can all be engaged.

Intel's theory goes something like this. Most apps use less than 8 cores. We'll have 8 big P cores to handle those apps. For the applications that are highly parallel we'll fill the rest of the die with E's to maximize parallel compute ability.

 

[Starjack]

I always wonder how the E-cores will perform, if the P-cores weren't included. Would they still give a respectable performance, at least handle some multitasking? I'm curious to know because I'm getting a next laptop with the Core i3-1215U (2 P cores /4 E cores (8 threads)) processor in it, in over a week's time.