Low Geekbench OpenCL scores - Anyone have any idea why?

[Dr. Stealth]

Hello All,

I need to tap into the expertise on this Forum....

I don't worry too much about benchmarks but they can be a good diagnostics tool. My other Geekbench scores in macOS High Sierra, CPU, CUDA & Metal seem to be around the same as others with near identical systems. My OpenCL score seems to be quite low compared to others. Not like a little bit but a lot, like 50% lower or more.

I thought it might be the GT-120 pulling the score down so I pulled the card out and get basically the same low score.

I'm not too worried about it as I mostly use CUDA but I'm sure wondering why they are so low and if there's anything I can do to bump them up to where they should be, or just don't worry about it.

Anyone have any ideas or input? Many Thanks !!

System Info:

Scores:

I know there's a difference between macOS support for OpenCL and Windows 10 but should Windows 10 OpenCL score, with the same hardware really be 230% faster? This is my score in Win 10.

 

[CWallace]

Apple continues to deprecate OpenCL as they try and push developers to Metal (2) so I would not be surprised if the Windows score was significantly higher.

 

[casperes1996]

Apple continues to deprecate OpenCL as they try and push developers to Metal (2) so I would not be surprised if the Windows score was significantly higher.

Not this much though - and the version of OpenCL on the Mac is still rather new and feature rich.

Then again - because that's what all newer Macs ship with, the optimisation work mainly goes into AMD designs.

 

[Dr. Stealth]

Thanks for your replies....

h9826790.... Waiting for your input.....

 

[AidenShaw]

Not this much though - and the version of OpenCL on the Mac is still rather new and feature rich.

Perhaps you should double check "what is the latest version of OpenCL" and "what is the latest version of OpenCL supported on Apple devices". You might be surprised.

Did you get the memo that Metal has replaced OpenCL? Apple really likes to adopt a new graphics/GPGPU framework every year or so.

 

[casperes1996]

Perhaps you should double check "what is the latest version of OpenCL" and "what is the latest version of OpenCL supported on Apple devices". You might be surprised.

Did you get the memo that Metal has replaced OpenCL? Apple really likes to adopt a new graphics/GPGPU framework every year or so.

2.2 and 1.2 respectively. Worth noting however that after 1.2 came 2.0 so it's not as big a gap as it seems. And it's nothing compared to the gaps that've been with OpenGL even before that got axed. From Snow Leopard all the way up to what? Mavericks? We just had something like 3.2 and then a massive leap up to 4.1.
i get that we're moving towards Metal, but I think OpenCL will stick around a while longer and get another revision bump before it gets the axe. Mainly because OpenCL offers the advantage that both CPU and GPU can run off of a shared code path in parallel. Apple's own software still also includes a fair amount of OpenCL implementations

 

[ticotoo]

Hello All,

I need to tap into the expertise on this Forum....

I don't worry too much about benchmarks but they can be a good diagnostics tool. My other Geekbench scores in macOS High Sierra, CPU, CUDA & Metal seem to be around the same as others with near identical systems. My OpenCL score seems to be quite low compared to others. Not like a little bit but a lot, like 50% lower or more.

I thought it might be the GT-120 pulling the score down so I pulled the card out and get basically the same low score.

I'm not too worried about it as I mostly use CUDA but I'm sure wondering why they are so low and if there's anything I can do to bump them up to where they should be, or just don't worry about it.

Anyone have any ideas or input? Many Thanks !!

System Info:
View attachment 743103

View attachment 743108

View attachment 743104

Scores:

View attachment 743106

I know there's a difference between macOS support for OpenCL and Windows 10 but should Windows 10 OpenCL score, with the same hardware really be 230% faster? This is my score in Win 10.

View attachment 743107

Your GTX1080Ti scores seem way to low. I'm getting ~17307 with my Mac Pro. I'm attaching runs from several of my tests. Notice little difference in CUDA scores on my 2014 15 " MBP using a egpu GTX 1080 and my MacPro. (this was done several months ago-can't directly compare)
I had tried to run the openCL benchmark earlier but Geekbench crashed. It wasn't until the more recent released versions of Geekbench that I could run the openCl benchmarks without crashing

. When I get a chance, after the holidays, I'll pull my GTX1080 out of my MacPro and try out the latest versions of Geekbench, High Sierra & the Nvidia drivers.

 

[flowrider]

Yep, way too low. I just ran the test with my GTX 1080.

Lou

 

[h9826790]

Thanks for your replies....

h9826790.... Waiting for your input.....

Are you comparing the same GPU? It looks like you have two 1080Ti. And each of them should able to achieve around 200000 score.

I am wondering if you are comparing the 1080Ti in slot 1 (x16 in Windows) to a 1080Ti in slot 3 (x4 in MacOS).

But even the bandwidth is limited to PCIe 2.0 X4, that score still looks too low to me.

Since the score in Windows is normal. If both of your 1080Ti experience the same issue. Then I will assume it’s the software issue. As you said, both cards works OK in almost all other situation. If that’s true. Then the biggest suspect will be Geekbench (Mac version) itself. The history proved that Geekbench can be buggy (e.g. crash in some specific GPU test).

If really want to find out what’s the exact bug. The very first thing I will try is only run a single 1080Ti, and see if the score back to normal. If yes, then I will suspect Geekbench (Mac version) may not able to handle multiple Pascal card properly in their OpenCL test.

 

[Synchro3]

Try macOS 10.12.6, maybe you get better results.

OpenCL in Windows isn't so much faster.

My GTX 1080 Ti

macOS 10.12.6:


Windows 10:

 

[Dr. Stealth]

I'll run some more tests in both Sierra and High Sierra and report back....

 

[Dr. Stealth]

Okay, I had a little time today to run a fresh series of Geekbench tests in both Sierra and High Sierra. Both are new, clean installs each on thair own SSD.

All tests were performed today with Geekbench 4.2.0. I have identicle 1080 TI's in both slot 1 and slot 4.

I see all scores making pretty good sense except for OpenCL.

I don't know if it matters at all but my display is plugged into the card in slot 1. This is the only thing I can think of that my be dropping the OpenCL score of the card in slot 1.


macOS Sierra 10.12.6

CPU
25,104

CUDA
Slot-1 245,707
Slot-4 242,260

Metal
Slot-1 225,136
Slot-4 224,347

OpenCL
Slot-1 151,607
Slot-4 212,664



macOS High Sierra 10.13.2

CPU
25,265

CUDA
Slot-1 233,126
Slot-4 245,064

Metal
Slot-1 229,307
Slot-4 231,765

OpenCL
Slot-1 96,994
Slot-4 195,520


Just for Reference: Win 10 Pro 1709

OpenCL
Slot-1 211,201
Slot-4 212,498

Slot 1 & 4 nearly identical.

 

[joebclash]

Is apple purposely slowing down older mac pro? Sorry, just joking. Just had to say it...

 

[casperes1996]

Is apple purposely slowing down older mac pro? Sorry, just joking. Just had to say it...

Well, depending on what their plans are for default configs of their upcoming modular MP is, they may be focusing less on multi-GPU setups like in the nMP, so they may in fact, not exactly cripple, but at least not optimise as much for that setup going forward. Unless of course multi-GPU takes a swing and interplay between iGPUs and dGPUs becomes super big and the code path to use an iGPU and a dGPU together would also mean the two dGPUs of the nMP would run together.

 

[flowrider]

Okay, I had a little time today to run a fresh series of Geekbench tests in both Sierra and High Sierra. Both are new, clean installs each on thair own SSD.

First off, there seems to be an issue with where the commas go in your scores. Second, where is Slot-1 - on the top or on the bottom?

Lou

 

[Asgorath]

I don't know if it matters at all but my display is plugged into the card in slot 1. This is the only thing I can think of that my be dropping the OpenCL score of the card in slot 1.

That'll be it, for sure. If OpenCL can run on the GPU without interruption from the display, it will generally be much faster.

 

[Dr. Stealth]

First off, there seems to be an issue with where the commas go in your scores. Second, where is Slot-1 - on the top or on the bottom?

Lou

Boy I blew the commas... Fixed.

Slot 1 is bottom x16

 

[Synchro3]

macOS Sierra 10.12.6
OpenCL
Slot-1 151,607
Slot-4 212,664

macOS High Sierra 10.13.2
OpenCL
Slot-1 96,994
Slot-4 195,520

Clearly an issue with High Sierra 10.13.2. Even if Sierra 10.12.6 is not optimal with two GTX 1080 Ti.

 

[flowrider]

^^^^My result in Sierra was a bit higher, but not by much.

Lou