Geekbench 4 is out. CPU and GPU..

[AidenShaw]

Just checked out the GB4 results...
I think something is way off with these results. In multicore, a 5820k does perform the same as a Xeon E5-2650v3 20 core CPU? Really?
https://browser.geekbench.com/v4/cpu/multicore?page=2

Makes perfect sense - see http://ark.intel.com/compare/82932,81705

The 2650 v3 is a 2.3 GHz 10 core processor, the 5820K is a 6 core 3.3 GHz processor.

 

[koyoot]

No it doesn't. What does makes sense is that the application reads it from 1 CPU, and there are two in the system.

Rest is completely understandable from this point of view.

 

[JronMasteR]

Makes perfect sense - see http://ark.intel.com/compare/82932,81705

The 2650 v3 is a 2.3 GHz 10 core processor, the 5820K is a 6 core 3.3 GHz processor.

Oh well ok you are right. But the result I am talking about is a 2 CPU system, so dual 2650v3's... 20 cores vs. 6 cores... So it seems that GB4 does not work correctly on dual cpu systems...
https://browser.geekbench.com/v4/cpu/119

 

[AndreeOnline]

…for Photoshop, should I pay attention to single-core performance or multi-core performance?

Single core.

 

[jerwin]

Just checked out the GB4 results...
I think something is way off with these results. In multicore, a 5820k does perform the same as a Xeon E5-2650v3 20 core CPU? Really?
https://browser.geekbench.com/v4/cpu/multicore?page=2

I looked at the raw data from my machine (a fairly bog standard 4 core machine without hyperthreading), and computed the throughput ratio.

Looks like the most parallel task was SFFT: 10.8 GFlops vs 42 GFlops. (3.8x speedup). Of the computational tasks, the most serial was LLVM: 357 ips vs 740.7 ips (2.1x speedup)

Looking at the 20 core beast, SFFT was 6.99 GFlops vs 150.3 GFlops (21.5x speedup) LLVM was 199.9 instructions vs 648.3 (3.24x).

Hyperthreading helps the tiniest bit. Take this imac with a 6700K

SFFT speedup is 4.07x; LLVM is 2.07x.

Of course, there is speedstep to consider-- 20 normally clocked cores are not going to be twenty times faster than a single overclocked core-- so maybe hyperthreading does make up some of the difference, and a little bit more.

 

[JimGoshorn]

Hope this questions is not too stupid: for Photoshop, should I pay attention to single-core performance or multi-core performance? Based on the Geekbench 4 Mac performance list, my machine (mid 2010 MP5,1, 2.93 GHz, 12 cores) is still quite good in multi-core performance, but quite lousy in single-core. I wonder if I should expect a major boost in Photoshop performance if I switch to MP6,1 or MP7,1.

Here is an article on Photoshop and it's multi-core performance:

https://www.pugetsystems.com/labs/articles/Adobe-Photoshop-CC-Multi-Core-Performance-625/

They are a custom builder of PC's but the info should still apply

 

[Synchro3]

Mac Pro 4,1 W3690 3.46 GHz 6 Core
Single: 3092
Multi: 13295


GTX 770
Compute CUDA: 78438


https://browser.geekbench.com/v4/cpu/36717

GTX 770
Compute OpenCL: 47176


Mac Mini 2012 2.3 GHz Quad Core
Single: 3188
Multi: 10321

https://browser.geekbench.com/v4/cpu/39287

 

[flowrider]

The GPU compute test quit before finishing.

GPU Test won't work for me either in Sierra. Haven't tried it yet in El Cap.

Lou

 

[pastrychef]

Based on what I see on their support forum, there seems to be quite a few little issues with Geekbench 4. Hopefully, they will get resolved quickly.

 

[Derpage]

It hurts that the fastest possible cMP (3.46 GHz) gets spanked in Single Core by anything remotely recent - even lowish-end mobile chips

I haven't tested this on any of my macs. I'm on an old pc with a 4820k in it and i'm still getting 4105 single and 15121 multi... I'm destroying these machines with a computer I built in 2013 and do not maintain or upgrade.

 

[thornslack]

Results for my dual x5677 (8-core 3.46) with GTX 780 and (Maxwell) Titan X. Too lazy to swap drives again to test the Titan under OSX. Windows ahead by a mile with GPU as usual. CPU scaling does seem a little wonky. Yesterday the compute test wouldn't complete in OSX, today it did. I think Geekbench has some kinks to iron out.


Win 10 CPU: https://browser.geekbench.com/v4/cpu/5126
+ Single-Core: 2855 Multi-Core: 16331

Win 10 Compute (m)TX: https://browser.geekbench.com/v4/compute/2271
+API score: 135903

Win 10 Compute 780: https://browser.geekbench.com/v4/compute/3322
+API score: 68848

OSX 10.11 CPU: https://browser.geekbench.com/v4/cpu/7754
+Single-Core: 3010 Multi-Core: 16302

OSX Compute 780: https://browser.geekbench.com/v4/compute/9602
+API Score: 58605

 

[Antoni Nygaard]

Mac Pro 2,1 with El Capitan, 10GB Ram. AMD 6870

 

[dfritchie]

Here's my Skylake Hackintosh scores:

 

[AidenShaw]

Results for my dual x5677 (8-core 3.46) with GTX 780 and (Maxwell) Titan X. Too lazy to swap drives again to test the Titan under OSX. Windows ahead by a mile with GPU as usual. CPU scaling does seem a little wonky. Yesterday the compute test wouldn't complete in OSX, today it did. I think Geekbench has some kinks to iron out.


Win 10 CPU: https://browser.geekbench.com/v4/cpu/5126
+ Single-Core: 2855 Multi-Core: 16331

Win 10 Compute (m)TX: https://browser.geekbench.com/v4/compute/2271
+API score: 135903

Win 10 Compute 780: https://browser.geekbench.com/v4/compute/3322
+API score: 68848

OSX 10.11 CPU: https://browser.geekbench.com/v4/cpu/7754
+Single-Core: 3010 Multi-Core: 16302

OSX Compute 780: https://browser.geekbench.com/v4/compute/9602
+API Score: 58605

 

[fastlanephil]

Mac Pro 2009 5,1 2X 3.46 GHz 6-core. 32GB 1333 memory

 

[thornslack]

Too bad there's no way to overclock the cMPs. The X5xxx Xeons were quite capable I think on other X58 motherboards. A 12-core cMP at 4.0+ GHz would be so cool.

 

[SoyCapitanSoyCapitan]

Too bad there's no way to overclock the cMPs. The X5xxx Xeons were quite capable I think on other X58 motherboards. A 12-core cMP at 4.0+ GHz would be so cool.

The mobo controller chip isn't well cooled and already prone to high temps. A tiny sometimes flaky heatsink and it is blocked from the system fans by the two CPU fans. PC mobos don't suffer this design.

 

[Amethyst]

As geekbench longtime user. I very sure that geekbench 4 significantly incorrected.

1) I think it no way, that 4.0 ghz skylake can be 20% faster than same 4.0 ghz haswell.
2) it no way that 4.0 ghz skylake score nearly to 12 core 3.46 ghz westmere.
3) it no way that single 7950 tahiti scored are same as dual d700 tahiti.
4) it no way that snapdragon 820 score are lower than exynos 7420.

So, from all early result, tell me that gb4 are totally nonsense, and primatelabs need to fix it now.

 

[ActionableMango]

5,1 Mac Pro
W3690 6-Core 3.46Ghz
Nvidia GTX980

     El Cap  Win10
Single: 3029   2832
Multi:  12023  12056
OpenCL: DNF   116457
CUDA:  129040  125751

It surprises me that El Cap is ahead of Win 10 on a couple of the results.

 

[Baby12]

iPhone 6 result

 

[scott.n]

Hackintosh
Haswell i7-4790K overclocked to 4.3 GHz
32GB RAM at 2133 MHz
GTX 970

CPU: 5027 / 16285

Compute test did not run.

 

[ActionableMango]

Yesterday the compute test wouldn't complete in OSX, today it did.

So far everyone reporting that Compute fails in OS X has a Maxwell card, so either the test doesn't work with Maxwell cards or it doesn't work with the Nvidia web drivers that Maxwell cards need.

But I see you have both Kepler and Maxwell cards. When the Compute test completed, which card and drivers were you using?

------

P.S. It is a little bit confusing that you have exactly the same avatar as AidenShaw, in the same thread, and quoting each other.

 

[SoyCapitanSoyCapitan]

5,1 Mac Pro
W3690 6-Core 3.46Ghz
Nvidia GTX980

     El Cap  Win10
Single: 3029   2832
Multi:  12023  12056
OpenCL: DNF   116457
CUDA:  129040  125751

It surprises me that El Cap is ahead of Win 10 on a couple of the results.

Geekbench was originally a Mac only app and always was showing a slight higher result in some tests (and Windows winning in others), but if you look at the difference in terms of percentage then it's within a margin of error.
[doublepost=1472751569][/doublepost]

As geekbench longtime user. I very sure that geekbench 4 significantly incorrected.

1) I think it no way, that 4.0 ghz skylake can be 20% faster than same 4.0 ghz haswell.
2) it no way that 4.0 ghz skylake score nearly to 12 core 3.46 ghz westmere.
3) it no way that single 7950 tahiti scored are same as dual d700 tahiti.
4) it no way that snapdragon 820 score are lower than exynos 7420.

So, from all early result, tell me that gb4 are totally nonsense, and primatelabs need to fix it now.

Maybe, but 6700K was killing X5690 on many real world tests I published here.

 

[thornslack]

So far everyone reporting that Compute fails in OS X has a Maxwell card, so either the test doesn't work with Maxwell cards or it doesn't work with the Nvidia web drivers that Maxwell cards need.

But I see you have both Kepler and Maxwell cards. When the Compute test completed, which card and drivers were you using?

My 780 is flashed so I keep it around as my EFI enabled option. Initially, compute test failed at the start of the histogram equalization test. Next day it was working. I didn't test the maxwell card in 10.11 because my snarky SSD/PCIe 6G and Win10 don't play nice with my soft raid OSX volume on the back plane

 

[AidenShaw]

Too bad there's no way to overclock the cMPs. The X5xxx Xeons were quite capable I think on other X58 motherboards. A 12-core cMP at 4.0+ GHz would be so cool.

Great avatar - everyone should try it!