Geekbench 6 results and discussions

[senttoschool]

Post Apple Silicon Geekbench 6 results or just any results from AMD/Intel/Qualcomm/other CPUs here.

Download: https://www.geekbench.com/download/mac/

• ST scores are now normalized to an Intel i7-12700 (2500 score). This means you need 5000 to double the ST performance of an Intel 12700.
• CPU and GPU GB6 results for an iPhone 13 Pro Max
• Generally, GB6 has bigger datasets such as larger photos and longer running tests than GB5 so bursting is less important.
• MT testing has been changed from each core having discrete tasks to cores working together on a single task. This reflects more real world and would capture interactions between cores such as core-to-core latency.
• Optimizes for NEON on ARM CPUs and AVX on x86 CPUs
• Tip: If you sign up for a free GB account, and append ".gb6" to the end of the results page URL, you can see the full detailed JSON with CPU frequencies. This is useful because there are a lot of suspicious high scores and you can check if the tester overclocked their CPUs. For example, this is an Intel 13900KF overclocked from 5.8Ghz to 6.18Ghz: https://browser.geekbench.com/v6/cpu/2541.gb6

 

[dmccloud]

I just ran GB6 twice - once with low power mode turned on and once without it enabled.

With Low Power Mode enabled:

Without Low Power Mode enabled:

That's a sizeable performance delta between the two modes. I'll have to post results for my AMD-based gaming rig later tonight.

 

[senttoschool]

My M1 Pro MBP 16":

 

[dmccloud]

The Metal score is surprising to me...

 

[Mike Boreham]

M2 MacBook Air:

 

[NotTooLate]

From some Anandtech browsing , looks like we will not be getting comparable results for the really multithreaded systems , the AMD Milan (server parts) are suddenly worse then desktop parts , looks like GB6 will be targeting lower thread count.
GB5 M2 vs Ulta is 3x or so , while GB6 is 2X I believe.

Lets see how things develop in the next few weeks , maybe the GB folks will tune it to support the bigger systems better or they will keep it consumer level for MT workloads.

 

[leman]

Looks like GB6 has fixed the problem with Apple GPUs. The M1 Max compute scores are now similar to the desktop RTX 3060, as expected.

 

[thenewperson]

Looks like GB6 has fixed the problem with Apple GPUs. The M1 Max compute scores are now similar to the desktop RTX 3060, as expected.

Are they now more comparable between each other? I remember GB5 you couldn’t really compare them.

 

[leman]

Are they now more comparable between each other? I remember GB5 you couldn’t really compare them.

Why not? It's the time to finish running a set of tasks. Why wouldn't GB5 results be comparable? The main problem with GB5 compute was that Apple GPUs need a considerable warmup period (somewhere around 1/5 of a second of I remember correctly) to trigger the high-performance mode and the tests were too short, so the GPU was running at somewhere between 50-70% of it's potential performance.

 

[senttoschool]

AMD Milan (server parts) are suddenly worse then desktop parts

That's expected. Most Epyc CPUs should be slower than Ryzen CPUs at consumer/prosumer tasks - even multithreaded. This is because 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" but the results are meaningless. Buyers of Epyc aren't looking at Geekbench scores, I can tell you that.

 

[leman]

That's expected. Most Epyc CPUs should be slower than Ryzen CPUs at consumer/prosumer tasks - even multithreaded. This is because 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".

Yep. And I think that what makes GB6 great. It is the first benchmark that really focuses on how applications actually work in desktop/workstation use. Previously, multicore performance was estimated by launching the same task across all cores. Every core was doing the same work. But that's not what we actually do. I am not compressing the same file 16 times. I want to use my 16 cores to compress one file faster. That's exactly what GB6 is trying to measure.

Servers are designed for something else entirely — running multiple completely independent tasks from different processes and even from different users. Of course, there will always be embarrassingly parallel algorithms where server CPU are a great fit (CPU raytracing, MCMC etc.). GB6 includes those types of applications as well. But overall, it's a much healthier mix of workloads that better reflect how we use computers.

I think would be great if one could additionally filter these results by benchmark, like having two different multicore scores — one for more inherently parallel applications and one for more complex cooperative ones. But you know, one step at a time.

 

[3Rock]

16" M2 Max 64gb 1Tb 38c

High Power and Low Power Readings

You guys might be interested in this YouTuber on the new Geekbench scoring

h

 

[senttoschool]

16" M2 Max 64gb 1Tb 38c

High Power and Low Power Readings

You guys might be interested in this YouTuber on the new Geekbench scoring

h

I generally like his videos but he's confused here. The scores are normalized to an Intel 12700.

Also, why would he want to compare GB5 results to GB6 results? Makes no sense.

 

[thenewperson]

Hmmm... I actually wonder how these were achieved.

 

[leman]

View attachment 2159003
Hmmm... I actually wonder how these were achieved.

I noticed those too, according to the JSON file it’s a mystery chip running at 40ghz or something ^^ wouldn’t pay much attention to it

 

[Juicy Box]

2023 M2 Pro Mac Mini, 12c CPU, 19c GPU, 32GB RAM:

 

[Adreitz]

View attachment 2159003
Hmmm... I actually wonder how these were achieved.

Doesn't surprise me. The GB result browser has been awful for years, so it wouldn't be unusual for them to not bother confirming any submitted results. I suspect most of the results on the first few pages of any GB listing are fake.

 

[NotTooLate]

That's expected. Most Epyc CPUs should be slower than Ryzen CPUs at consumer/prosumer tasks - even multithreaded. This is because 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" but the results are meaningless. Buyers of Epyc aren't looking at Geekbench scores, I can tell you that.

Of course , the intent of my post was to be careful of even running it on the Ultra for example , as you can see it went from 3x better MT then M2 to 2x only , the Milan example is the extreme side of the new GB6 limitations that's all.
No one Is running GB for anything important , its for internet glory and reviewers to get easy "this is 20% faster then previous CPU" by clicking a button and sounding smart , if you have a SW you work with you benchmark using that program.

 

[xraydoc]

14" MacBook Pro M1 Pro (10c/16c, 16GB RAM):

 

[WilliApple]

The complaints I have is that:
1. There is no history in the iOS version
2. There is no standalone macOS version that cost $10.

Once those features come, I will make the switch to 6.

But I did run one on my M1 Air and the scores were higher.

 

[xraydoc]

Hmmm... I actually wonder how these were achieved.

Fake most likely.

 

[leman]

Of course , the intent of my post was to be careful of even running it on the Ultra for example , as you can see it went from 3x better MT then M2 to 2x only , the Milan example is the extreme side of the new GB6 limitations that's all.

It's not a limitation, it's just that scaling multicore workloads is hard. For example, this is a test of building a very large codebase (thousands and thoudanss of source files) on the largest CPUs money can buy:

Timed LLVM Compilation Benchmark - OpenBenchmarking.org

openbenchmarking.org

Note how you quickly start experiencing diminishing returns when increasing the number of cores.

What I really like about GB6 is that it really puts a very realistic perspective on multicore results. Most benchmarks are very naive in this regards. This is why Intel can add 16x slow cores to their processors and claim much better MT performance — these benchmarks will agree. But with GB6 actually testing the ability of the cores to cooperate towards a single task the problem with this approach become quickly apparent.

No one Is running GB for anything important , its for internet glory and reviewers to get easy "this is 20% faster then previous CPU" by clicking a button and sounding smart , if you have a SW you work with you benchmark using that program.

That for sure, always benchmark for your use case. But the new methodology of GB6 gives a nice overall estimate of different types of general-purpose workloads.

 

[jeanlain]

What I really like about GB6 is that it really puts a very realistic perspective on multicore results. Most benchmarks are very naive in this regards. This is why Intel can add 16x slow cores to their processors and claim much better MT performance — these benchmarks will agree. But with GB6 actually testing the ability of the cores to cooperate towards a single task the problem with this approach become quickly apparent.

Does this favour Apple SoCs?

EDIT: The M1 Ultra score is 7.4 higher in MT compare to ST. The ratio is about the same for the intel 13900k, despite a much higher number of threads (and more cores).
I wonder how much of this is due to better multithreading for the M1 ultra or the fact that the clock speed is less reduced in MT tasks.

 

[leman]

Does this favour Apple SoCs?

Hard to say. From a cursory glance it seems that each P-cluster (4 P cores) adds about 4K points, or 50% of M1 MT score. So there is definitely some inefficiency, but without detailed comparison I don’t immediately see how it relates to other devices.

It’s important to note however that a P-cluster in Apples design shares L2 cache, so communication between cores in a cluster will be very efficient. Exchanging data between clusters is probably much slower (which is likely what we see here). I wonder how this compares to AMD desktop designs which have similar topology.

EDIT: The M1 Ultra score is 7.4 higher in MT compare to ST. The ratio is about the same for the intel 13900k, despite a much higher number of threads (and more cores).
I wonder how much of this is due to better multithreading for the M1 ultra or the fact that the clock speed is less reduced in MT tasks.

It’s really hard to make sence of since the Intel chip has 16 lower-performance cores and 8 P-cores, where Ultra has 20 P-cores.

 

[Juicy Box]

2023 M2 Pro Mac Mini, 12c CPU, 19c GPU, 32GB RAM:

View attachment 2159044

View attachment 2159045

Here is my 2020 M1 Mac Mini, 16GB RAM to compare to the M2 Pro I posted earlier:




Quick GB6 comparison between the M1 Mac Mini, 16GB RAM to the M2 Pro Mac Mini, 12c/19c, 32GB RAM:

Single core - The 12c M2 Pro CPU is almost exactly 15% faster than the M1 CPU
Multi core - The 12c M2 Pro CPU is almost about 72% faster than the M1 CPU
Metal GPU - The 19c M2 Pro GPU is almost about 208% faster than the M1 GPU <-- Impressive

While I have been impressed overall with my M2 Pro so far, the GPU performance increase over the M1 is huge. I have only tested it on WoW so far, but the results are impressive, with over three times the frame rate at max graphic settings in Dragon Flight areas.