Geekbench 5 vs Geekbench 4

Eug

Lifer
Mar 11, 2000
23,586
1,000
126
Geekbench 5 just officially launched. So, is Geekbench 5 a better cross-platform test than Geekbench 4?

CPU Benchmark
The Geekbench 5 CPU Benchmark includes new benchmark tests that model the challenges your system faces when running the latest applications. These tests use cutting-edge technologies, including machine learning, augmented reality, and computational photography.

Geekbench 5 also increases the memory footprint of existing workloads to more accurately account for the effect memory performance has on CPU performance.

Finally, the Geekbench 5 CPU Benchmark includes new modes of multi-threaded benchmarks, allowing threads to work co-operatively on one problem rather than separately on different problems. With the addition of different threading models, Geekbench 5 better captures the performance of different multi-threaded applications on personal computing devices.

GPU Compute Benchmark
Geekbench 5 includes several improvements to the GPU Compute Benchmark.


The most exciting change is that the Compute Benchmark now supports Vulkan, along with CUDA, Metal, and OpenCL. Vulkan is the next-generation cross-platform graphics and compute API. Vulkan Compute Benchmarks are available on Android, Windows, and Linux.

The Compute Benchmark also includes new benchmark tests that model algorithms that are GPU accelerated in modern applications. These include computer vision tasks such as Stereo Matching, and augmented reality tasks such as Feature Matching.
 
  • Like
Reactions: Tlh97 and amd6502

amd6502

Senior member
Apr 21, 2017
971
360
136
I'm not sure if GB5 will break from the usual GB, but I think it would be nice if a benchmark could test for mixed multithread loads.

Traditionally to get a MT score in GB (or most/all benchmarks I've seen) you have it compute a problem n-times in parallel. And typically this problem is either floats heay or int heavy, in order to test FPU and CPU (ints) respectively.

I'd more like a bench to resemble real life loads.

How does it compute a stream of n-threads of floaty problems, while simultaneously computing a stream of m-threads of inty or mixed problems. Or some combination of such.

Secondly, how does the memory system perform while performing 1 thread, 3 threads, and 7 threads.
 

lopri

Elite Member
Jul 27, 2002
13,209
594
126
I've just run it on my 3700X.

ST 1324
MT 10227


Linux version does not perform much better this time. Heh.

ST 1339
MT 9717

 

Nothingness

Platinum Member
Jul 3, 2013
2,405
736
136
I've just run it on my 3700X.

ST 1324
MT 10227


Linux version does not perform much better this time. Heh.

ST 1339
MT 9717

Yeah this time they used the same compiler on Android, Linux and Windows. More information in this document: https://www.geekbench.com/doc/geekbench5-cpu-workloads.pdf
 

DrMrLordX

Lifer
Apr 27, 2000
21,620
10,830
136
Looks interesting enough. The weighting scale does not separately score the memory subsystem. I like that improvement over GB4.

My system:


Easy enough to run @ 4425 MHz static. I had a better score earlier but oh well.
 
  • Like
Reactions: amd6502

lopri

Elite Member
Jul 27, 2002
13,209
594
126
Did you guys know that you can check the clock frequencies at your CPU was run by typing in .gb5 (or .gb4 for Geekbench 4) at the end of the URL of the score page? I recently found out about it and thought it's cool. Like this:

Code:
https://browser.geekbench.com/v5/cpu/119259
https://browser.geekbench.com/v5/cpu/119259.gb5

The second address gives you the details page.
 
  • Like
Reactions: Eug

lopri

Elite Member
Jul 27, 2002
13,209
594
126
Looks interesting enough. The weighting scale does not separately score the memory subsystem. I like that improvement over GB4.
I kind of liked the memory sub-scores given, though. Maybe it should not be factored into the overall score, but I think having them there for additional data points might have been a better idea.
 

DrMrLordX

Lifer
Apr 27, 2000
21,620
10,830
136
I kind of liked the memory sub-scores given, though. Maybe it should not be factored into the overall score, but I think having them there for additional data points might have been a better idea.

Yeah that would be fine. It sure throws off the scores in GB4 though.
 

Eug

Lifer
Mar 11, 2000
23,586
1,000
126
This is a friggin phone!

Geekbench 4: 5472 / 13769
Geekbench 5: 1324 / 3394

I believe it's the iPhone 11 Pro, judging by the iPhone12,3 designation and the motherboard identifier.
 

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
Sounds like Apple claims about 20% were spot on. They achieve those gains everywhere, in Integer and FP performance. This 2.66Ghz CPU now has same single thread performance as Zen2/Skylake ~4.5Ghz.
 

nicalandia

Diamond Member
Jan 10, 2019
3,330
5,281
136
Sounds like Apple claims about 20% were spot on. They achieve those gains everywhere, in Integer and FP performance. This 2.66Ghz CPU now has same single thread performance as Zen2/Skylake ~4.5Ghz.
Could a custom chip for servers made from this uArch CPU defeat Epyc Rome with its 128/256 2P server?
 

JoeRambo

Golden Member
Jun 13, 2013
1,814
2,105
136
Could a custom chip for servers made from this uArch CPU defeat Epyc Rome with its 128/256 2P server?

Servers are more than just great ST performance. Need coherency between cores and sockets, interconnects, RAS features etc. Probably would need different cache and "uncore" architecture, cause right now it is sharing L2 cache between cores and obviously that gives great ST performance when only one is active. Can't really extend this Core2 era model beyond maybe 3-4 cores.

But obviously with Apple's talent and their infinite resources it would be very doable for them technically to achieve superior server chips in several generations. The real questions are political and economical i think.