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:

 

[MS_AT]

There is a difference between SME/AVX-512 SME is not part of an individual core it is shared by the entire CPU while AVX-512/VNNI is per core and when comparing single core performance how can you use a shared processing unit. It's like one core is using the ALU of another core

It's an interesting argument. What about memory controller? In the end if the shared unit is a bottleneck then MT tests should prove it. After all GB has high quality MT tests, right?

 

[okoroezenwa]

I would think ARM's SME would be inside the core since they sell core IP.

that’s my assumption as well. Not sure what QC will do though, since they got a lot of Apple people so there’s a chance they will just do similar.

 

[511]

It's an interesting argument. What about memory controller? In the end if the shared unit is a bottleneck then MT tests should prove it. After all GB has high quality MT tests, right?

Well than if someone uses NPU in GB Single Core to boost Object Detection .
My Point is if SME is part of single core it's fine to be included in SC Score else big No

 

[Doug S]

I would think ARM's SME would be inside the core since they sell core IP.

If it is in the core it will likely not be as well resourced as Apple's - i.e. it might be as fast or faster in MT tests with all the cores' small SME units working together, but not as fast in ST.

 

[igor_kavinski]

If suppose the SME instructions in the GB6 source code are handwritten while the AVX-512 instructions are just code vectorization by using a compiler switch, that would be unfair because compilers aren't good at spitting out performant AVX-512 code.

 

[Hitman928]

You mean like when Zen4 got 242.1% higher uplift in Object Detection over Zen3 via AVX upgrade?

That's higher than the 227.9% that M4 Max got over M3 Max.

I didn't see you complain then.

Zen3 vs Zen4: https://browser.geekbench.com/v6/cpu/compare/13703871?baseline=13703302

M3 Max vs M4 Max: https://browser.geekbench.com/v6/cpu/compare/13690938?baseline=13701855

View attachment 129833

https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

Didn't someone here run AVX512 on and off comparisons for GB6 on Zen 4 and find almost no difference in performance?

 

[511]

You mean like when Zen4 got 242.1% higher uplift in Object Detection over Zen3 via AVX upgrade?

That's higher than the 227.9% that M4 Max got over M3 Max.

I didn't see you complain then.

Zen3 vs Zen4: https://browser.geekbench.com/v6/cpu/compare/13703871?baseline=13703302

M3 Max vs M4 Max: https://browser.geekbench.com/v6/cpu/compare/13690938?baseline=13701855

View attachment 129833

https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

Look at the multi score as well for AMD Zen 4 vs Zen 3 and M4 vs M3 the gain is significantly reduces

 

[poke01]

there certainly is a difference when avx512 is used

Micro-Star International Co., Ltd. MS-7E47 vs MacBook Pro (14-inch, Nov 2023) - Geekbench

 

[Hitman928]

there certainly is a difference when avx512 is used
View attachment 129840

Micro-Star International Co., Ltd. MS-7E47 vs MacBook Pro (14-inch, Nov 2023) - Geekbench

AVX512 vs Neon?

 

[mikegg]

Look at the multi score as well for AMD Zen 4 vs Zen 3 and M4 vs M3 the gain is significantly reduces

OK so the MT score is lower than it should be. Who cares? Just focus on overall ST and overall MT scores.

 

[511]

With the way things are going with Geekbench Cinebench R23/24 seems a better benchmark.

 

[igor_kavinski]

At least GB6 is testing multiple workloads. What prevalent and widely desired real world use case is cindymama testing exactly?

I posted a thread for Upscayle. It's a better benchmark than cindinga. At least you can use the application to upscale tiny pics from the internet and go, woah! That's so cool!

 

[igor_kavinski]

Double post thanks to bots overwhelming this site.

It's easier to post than edit a post during such times.

 

[511]

Wow Igor at least you are making multiple copies of your messages unlike Geekbench who doesn't even make 2 copies of the benchmark to be run on 64 Cores

 

[igor_kavinski]

unlike Geekbench who doesn't even make 2 copies of the benchmark to be run on 64 Cores

Pfftt...you can do that with Powershell script. Then it's your job to make sense of the two different benchmark URLs.

 

[igor_kavinski]

New iPhone chip beating M4 Max in three ST subtests: https://browser.geekbench.com/v6/cpu/compare/13726766?baseline=13736665

 

[511]

New iPhone chip beating M4 Max in three ST subtests: https://browser.geekbench.com/v6/cpu/compare/13726766?baseline=13736665

~1% IPC Improvement
M4 Max -> 906.4(4079@4.5)
iPhone -> 920.8 (3895@4.2)

 

[Doug S]

~1% IPC Improvement
M4 Max -> 906.4(4079@4.5)
iPhone -> 920.8 (3895@4.2)

Given that object detection is slightly faster on iPhone despite the lower clock speed, I think A19P must have faster L2 and/or bandwidth (or faster LPDDR5X?) than M4 Max. I really doubt they made the SME unit bigger, but just a tiny tiny bit bigger.

 

[poke01]

iPhone18,1 - Geekbench

Benchmark results for an iPhone18,1 with an ARM processor.

browser.geekbench.com

New score for A19 Pro

 

[igor_kavinski]

vs. 268V: https://browser.geekbench.com/v6/cpu/compare/13747615?baseline=13668886

Can't overcome the lack of threads in half of the MT subtests. Still pretty potent if this ends up in a $799 MBA laptop.