Question Geekbench 6 released and calibrated against Core i7-12700

[igor_kavinski]

Geekbench 6 - Geekbench Blog

www.geekbench.com

Weird choice of baseline CPU and even weird is that the baseline score is 2500.

i7-12700 does hardly 2000 in GB5 with the fastest DDR5.

EDIT: Gideon's browser extension/add-on: https://forums.anandtech.com/thread...d-against-core-i7-12700.2610597/post-41452115

Helps you make sense of the results if you are left scratching your head about why a certain test is fast on a certain CPU.

Super useful!

Sample screenshot:

 

[poke01]

Well, choosing 12700 and removing AMD's advantage just means they are shamelessly pro-Intel.

just to note. Intel was always chosen as the baseline. No AMD, Apple or QC chip was ever chosen.

 

[poke01]

Coz heavy multithreaded would favor AMD!

WOW. Are they being paid off by Intel?

Sigh. They are not being paid by anybody. GB6 is highly tuned to be CPU agnostic. Some say they are being paid by Intel or Apple. AVX512 is not used by ARM CPUs, thats why they removed it.

 

[igor_kavinski]

just to note. Intel was always chosen as the baseline. No AMD, Apple or QC chip was ever chosen.

True, but the previous baseline was a Core i3 with a baseline score of 1000. They didn't give any reasons for their new choice. It's like they WANTED to have the E-cores represented.

 

[igor_kavinski]

AVX512 is not used by ARM CPUs, thats why they removed it.

OK, that makes sense.

 

[poke01]

True, but the previous baseline was a Core i3 with a baseline score of 1000. They didn't give any reasons for their new choice. It's like they WANTED to have the E-cores represented.

I saw on Twitter that Primate Labs expect CPUs to be very powerful next few years. With Apple and AMD being very good they expect CPUs to improve dramactically. Once Arrow Lake comes out, the 12th gen i7 won't look as good.

 

[Markfw]

just to note. Intel was always chosen as the baseline. No AMD, Apple or QC chip was ever chosen.

And this is somehow fair ???? Right now you have AMD as king of servers, HEDT, workstation, and right on top of desktop. They might even own laptop/mobile.
And to remove avx-512 ? who cares if arm or now intel do not support it. Its still a major player.

 

[poke01]

And to remove avx-512 ? who cares if arm or now intel do not support it. Its still a major player.

Its really not... That would give AMD an advantage. It would not be useful to compare with ARM or Intel.

They might even own laptop/mobile

After seeing M2 Pro/Max and thats still on 5nm. I believe Apple is ahead for now.

And this is somehow fair ???? Right now you have AMD as king of servers, HEDT, workstation, and right on top of desktop.

yeah it is strange they would use Intel but I guess they wanted a popular CPU. idk.

 

[Hitman928]

AVX512 is not used by ARM CPUs, thats why they removed it.

So why did they include it in previous versions when only Intel supported it? GB6 also still supports AVX2 which ARM also doesn’t support. I’m not saying they were paid to remove it, but the decision making seems questionable.

 

[Markfw]

Its really not... That would give AMD an advantage. It would not be useful to compare with ARM or Intel.



After seeing M2 Pro/Max and thats still on 5nm. I believe Apple is ahead for now.


yeah it is strange they would use Intel but I guess they wanted a popular CPU. idk.

Regardless of avx-512. Not to properly support multicore CPU's across the board, invalidates most CPUs of today.

 

[poke01]

Regardless of avx-512. Not to properly support multicore CPU's across the board, invalidates most CPUs of today.

Yeah I agree but I never used Geekbench for MT. I have always used it for ST. For MT there are better applications like Blender.

 

[poke01]

So why did they include it in previous versions when only Intel supported it? GB6 also still supports AVX2 which ARM also doesn’t support. I’m not saying they were paid to remove it, but the decision making seems questionable.

Very good questions.

 

[igor_kavinski]

Very light benchmark, hardly any benefit enabling SMT on a 7950x.

There is a pretty tangible benefit in some tests like Clang and raytracing:

ASUS System Product Name vs ASUS System Product Name - Geekbench

Score is ruined by these:



@Carfax83, can you do a comparison of HT/non-HT to see if these tests also affect Intel as badly too?

 

[JoeRambo]

Score is ruined by these:

And equally helped by








What people fail to realize is that very few tasks have perfect scaling with cores, rendering is best known scaling perfectly ( Cinebenches and co ), compilation scales worse, but can still extract something out of each additional core. There are natural cases like ray tracing, where you can scale the number of rays being cast (calculated), but you can't use 128 cores to calculate each ray faster.

GB6 in my opinion makes perfect case for desktop and workstation case:
think about it this way - there is a difference between PDF rendering to quickly open and render some files on screen and a processing pipeline where you take some bunch of files and convert them to PDF using multiple cores.

Plenty of guys use PDF, very few batch convert and GB6 leans towards the users.

There's also acknowledgment that computing is changing, gone are the "render like" task as noone is actually using CPUs to render anymore, it's all about GPUs now.
Also gone are the retarded encryption acceleration checks. Made zero sense in desktop/workstation environment and was great joy of incompetent CPU developers.

 

[naukkis]

MT score seems to highly depend on ST score somehow, for most scores so far of current chips with 8 cores or (even significantly) more the MT score is about 6-8 times the ST score. The only exception I saw so far is the 64c ARM server posted on last page with a ratio of ~13.

That's actually how most workloads scale to MT. Main thread is dominant and there's a limit how much can be offloaded to child threads. ARM designs for phones are tuned preciously to that, one prime core to run main thread and somehow less powerful cores to offload main thread. Only very special case of works scale to unlimited number of threads - and for normal desktop/phone use cores beyond ~8 are just as beneficial as Geekbench6 shows them to be.

 

[mikegg]

And equally helped by
View attachment 76529

View attachment 76530

View attachment 76531



What people fail to realize is that very few tasks have perfect scaling with cores, rendering is best known scaling perfectly ( Cinebenches and co ), compilation scales worse, but can still extract something out of each additional core. There are natural cases like ray tracing, where you can scale the number of rays being cast (calculated), but you can't use 128 cores to calculate each ray faster.

GB6 in my opinion makes perfect case for desktop and workstation case:
think about it this way - there is a difference between PDF rendering to quickly open and render some files on screen and a processing pipeline where you take some bunch of files and convert them to PDF using multiple cores.

Plenty of guys use PDF, very few batch convert and GB6 leans towards the users.

There's also acknowledgment that computing is changing, gone are the "render like" task as noone is actually using CPUs to render anymore, it's all about GPUs now.
Also gone are the retarded encryption acceleration checks. Made zero sense in desktop/workstation environment and was great joy of incompetent CPU developers.

This is correct.

I expect a highly clocked Ryzen desktop CPU to beat any Epyc CPU in most desktop class applications. Geekbench is designed to measure mobile apps to desktop apps. It's not designed to measure server/cloud CPUs.

Consumer/prosumer applications do not utilize that many cores and Epyc's many cores do not make up for the difference caused by much lower clock speeds.

I don't think server CPUs should be using Geekbench at all. It's just not the right tool to measure server CPU performance. Once in a while, someone will run Geekbench with a 64/96-core Epyc for "fun".

 

[roger_k]

I applaud Poole for their new approach on estimating multicore performance. Nobody is running the same task replicated across every physical core. Measuring multi-core benchmark in this naive way essentially makes every task embarrassingly parallel and massively overestimates the real-world performance impact of many-core designs in the desktop/workstation space. GB6 seems to have a good mix of tasks which tend to be embarrassingly parallel (like ray tracing) as well as tasks that have complex dependencies (like compiling code). This is a step in the right direction.

It would be nice of the GB website showed multicore scaling for each of the task, this would allow one to evaluate the advantages and disadvantages of various CPU designs (like Intel's reliance on many small throughput cores vs. AMDs scalable symmetric cores) for cooperative task solving.

 

[mikegg]

Regardless of avx-512. Not to properly support multicore CPU's across the board, invalidates most CPUs of today.

Again, it's not the right way to look at it. GB6 is meant for mobile or desktop-class applications. These applications do not make use of many cores. And MT workloads generally do not scale unless it's something like CPU rendering or some cloud application.

At some point, having too many cores will decrease performance for most applications due to lower clock speeds and less efficient scheduling.

In my opinion, GB6 is a better representation of real-world applications than GB5.

I'm a software engineer.

 

[mikegg]

Not everything is a conspiracy theory to make AMD CPUs look worse.

 

[mikegg]

I applaud Poole for their new approach on estimating multicore performance. Nobody is running the same task replicated across every physical core. Measuring multi-core benchmark in this naive way essentially makes every task embarrassingly parallel and massively overestimates the real-world performance impact of many-core designs in the desktop/workstation space. GB6 seems to have a good mix of tasks which tend to be embarrassingly parallel (like ray tracing) as well as tasks that have complex dependencies (like compiling code). This is a step in the right direction.

It would be nice of the GB website showed multicore scaling for each of the task, this would allow one to evaluate the advantages and disadvantages of various CPU designs (like Intel's reliance on many small throughput cores vs. AMDs scalable symmetric cores) for cooperative task solving.

Agreed. If anything, the new MT test should make Intel look worse and AMD look better. Intel's strategy of "stuff as many little cores as possible to win at Cinebench" is much less effective when the cores actually have to work together on the same task, which is how most applications are designed.

GB6's MT scaling should be much closer to real-world application MT scaling than GB5's.

I think people here are shocked to finally find out that the vast majority of applications do not scale linearly at all with the number of cores. Maybe I shouldn't say shocked. Maybe I should say that most people here were delusional. They knew MT doesn't scale as Cinebench suggests but they still insist on buying ridiculous high-core CPUs anyway. They needed Cinebench and GB5 to validate that their purchase was necessary.

 

[roger_k]

Regardless of avx-512. Not to properly support multicore CPU's across the board, invalidates most CPUs of today.

Quite in contrary, this new approach is arguably more representative.

The way how multi-core benchmarks are usually done is by running the same task on all cores at the same time. But this is hardly an interesting measure. If I am converting my complex poster to PDF, I am not doing it 20 times simultaneously. I am doing it only once and I want all the cores in my machine to work together in oder to speed up the process. This is obviously a more complex task which is unlikely to get perfect multicore scaling. And that's exactly what GB6 is measuring. I would think is much more relevant to the end user.

P.S. I do think that excluding AVX-512 was a bit of a shame. If a CPU supports the technology and can perform faster on a task, it should be used. At the same time, I do understand the decision — AVX512 isn't really used all that much in desktop software, and basing benchmarks off it might create unrealistic expectations.

 

[Exist50]

Agreed. If anything, the new MT test should make Intel look worse and AMD look better. Intel's strategy of "stuff as many little cores as possible to win at Cinebench" is much less effective when the cores actually have to work together on the same task, which is how most applications are designed.

I think it should be pretty neutral, net for net. The same argument applies for SMT helping bump up AMD's throughput numbers. And this might potentially expose some cross-CCX dependencies.

But regardless of who comes out looking better, this is objectively more representative of the behavior real user programs exhibit, and so should be applauded as a good step forward. No one seriously in the market for, say, servers, was never using Geekbench to begin with, and even workstation type tasks have better test suites available (e.g. PugetBench).

 

[mikegg]

Here's my chance to bring back my thread on Cinebench on Reddit.

The great ex-Anandtech Andrei F chimed in.

Through marketing, first AMD, and now Intel, the x86 CPU industry has brainwashed people into using Cinebench as a general-purpose CPU benchmark despite the fact that it does not correlate with the applications 99% of people use. I think GB6's MT results are a significant improvement on both GB5 and Cinebench.

https://www.reddit.com/r/hardware/comments/pitid6

 

[igor_kavinski]

There is still a flaw inherent in GB6's testing methodology. It's not measuring multitasking efficiency which is what AMD excels at. They should run at least two different tests concurrently to see how the CPU handles multiple workloads thrown its way.

 

[Kocicak]

There is still a flaw inherent in GB6's testing methodology. It's not measuring multitasking efficiency which is what AMD excels at.

How do you know that? What benchmarks measure simultaneusly running different sorts of tasks?

 

[mikegg]

multitasking efficiency

Can you define this more?

There is still a flaw inherent in GB6's testing methodology. It's not measuring multitasking efficiency which is what AMD excels at. They should run at least two different tests concurrently to see how the CPU handles multiple workloads thrown its way.

Do we have to have benchmarks that are designed to favor AMD? It's never going to be 100% fair.

Take for example, Apple Silicon dedicates a lot of transistors to media encoders, decoders, neural engine, image processing. GB6 does not make use of any of these accelerators in the test but real world applications use them to speed up tasks. Is that fair? What if Apple decided to remove these accelerators and used all the transistors for more CPU L1/L2/L3 cache or just more CPU transistors? Now Apple Silicon's CPU would be much faster but the overall experience would be worse for users.