15-inch MBP Geekbench GPU results much worse on 10.12.3

[Eldar Gezalov]

Hey guys,

I just wanted to check if I am the only one who gets much worse Geekbench test results for GPU performance after updating to 10.12.3. Looks like the problem is only with 15-inch MacBook Pro, because I have also tested this on 13-inch tbMBP and the results are the same.

I have 15-inch MacBook Pro with 2.7/16/512/455 specs, and on 10.12.2 my compute score for Radeon Pro 455 was over 50000, while on 10.12.3 only about 40000.

Check out the links for compute results before and after:

Before: https://browser.geekbench.com/v4/compute/397429
After : https://browser.geekbench.com/v4/compute/400887

 

[Sanpete]

Quite a difference! Did you restart the machine between the tests, or could there have been some other factor that affected the tests? Things can slow down sometimes between restarts.

 

[Eldar Gezalov]

Quite a difference! Did you restart the machine between the tests, or could there have been some other factor that affected the tests? Things can slow down sometimes between restarts.

Yes, I did restart the machine many times, and I did a lot of tests. Every single time the similar output. It would be nice to see what results other people get with 15 MBP on 10.12.3.

 

[Photios]

Yes, I did restart the machine many times, and I did a lot of tests. Every single time the similar output. It would be nice to see what results other people get with 15 MBP on 10.12.3.

Can you tell that the discreet GPU is being used as opposed to the integrated GPU?

 

[Eldar Gezalov]

Can you tell that the discreet GPU is being used as opposed to the integrated GPU?

Yes, I checked that as well. I even manually set GPU to discreet, same result.

I checked the latest results on Geekbench website, and noticed that many 15-inch MBPs on 10.12.3 show the similar decrease in compute score.

Check it here: https://browser.geekbench.com/v4/compute/search?utf8=✓&q=MacBook+Pro+(15-inch+Late+2016)

 

[nicolipo]

wow that's true! my gpu is curbed too!
seems like apple wants to extend the battery-endurance at the expense of performance... ?!

 

[keysofanxiety]

Yes, I checked that as well. I even manually set GPU to discreet, same result.

I checked the latest results on Geekbench website, and noticed that many 15-inch MBPs on 10.12.3 show the similar decrease in compute score.

Check it here: https://browser.geekbench.com/v4/compute/search?utf8=✓&q=MacBook+Pro+(15-inch+Late+2016)

Really odd question, but could you try to plug in an external monitor and try the benchmarks again?

It seems almost unarguable that Apple have downclocked/throttled the GPU to extend battery life, however I'm not sure if that's permanent or if it would revert under stress. Granted, you'd think that benchmarking would count as GPU intensive but perhaps it would come into play with an external monitor plugged in and the laptop is charging at the same time.

I just really hope that people don't see a negative impact to their workflow as a result of this; fingers crossed that it's another case of benchmarks not reflecting the real-world scenario. I wonder if anybody has any before/after comparisons with gaming framerates or LPX/FCPX performance?

 

[agaskew]

GPU performance has perhaps been dialled down, to reduce the GPU corruption/problems seen by some.

 

[leman]

What about real-world OpenCL workflows? I have stopped trusting Geekbench long time ago. Its not only non-representative of real-world performance but also gets some basic benchmarks (like RAM bandwidth) completely wrong.

Running some 3D benchmarks didn't show any regression, in fact, the results have increased by around 3% (I wouldn't read much into it, probably random fluctuation).

 

[aevan]

I'm getting around 42000 now, but I ran it a few weeks ago and I got something similar too. I can't remember how much exactly, but it wasn't 50000 if I remember. OP, are you sure the first test was accurate (maybe it was some glitch)? Can anyone else confirm this?

Also, OP, did you run the tests attached to power? (you probably did, but asking just in case).

 

[Eldar Gezalov]

Really odd question, but could you try to plug in an external monitor and try the benchmarks again?

I did what you asked, same result.

What about real-world OpenCL workflows? I have stopped trusting Geekbench long time ago. Its not only non-representative of real-world performance but also gets some basic benchmarks (like RAM bandwidth) completely wrong.

This is my opinion right now. I didn't notice any drops in performance (although it is difficult for me to notice drops in performance, because my workflow does not include any GPU intensive tasks).

I'm getting around 42000 now, but I ran it a few weeks ago and I got something similar too. I can't remember how much exactly, but it wasn't 50000 if I remember. OP, are you sure the first test was accurate (maybe it was some glitch)? Can anyone else confirm this?

Also, OP, did you run the tests attached to power? (you probably did, but asking just in case).

Yes, I did run the tests attached to power, similar results. I am getting results between 39000 ~ 42000, so similar to your machine. Before the update, I was always getting something like 49000~50000. I am sure about it, because I have tested my machine a lot for a review for my blog.

Also, I checked the results of other users on Geekbench site, and it looks like the most machines with Radeon Pro 455 on 10.12.2 were all getting on average over 48000.
[doublepost=1485354846][/doublepost]

Really odd question, but could you try to plug in an external monitor and try the benchmarks again?

I did a test multiple times when connected to external monitor, and I actually got 44000 once. Without external monitor most of the times I get 37000 to 41000, so external monitor really makes some difference.

 

[MrGuder]

Do you think Apple would do something like this and not think the tech people wouldn't find out? Reducing performance to increase battery?

I think we need more testers to try this to see for sure.

 

[Sanpete]

If the update is affecting dGPU performance, it's not likely to be intended to affect battery life, since the kind of ordinary use tested and talked about for battery life doesn't involve the dGPU. It may well be something related to the graphics glitches some were seeing, as @agaskew suggested above. Will be interesting to see more data, and whether the graphics glitches lessen.

 

[jerryk]

If the update is affecting dGPU performance, it's not likely to be intended to affect battery life, since the kind of ordinary use tested and talked about for battery life doesn't involve the dGPU. It may well be something related to the graphics glitches some were seeing, as @agaskew suggested above. Will be interesting to see more data, and whether the graphics glitches lessen.

That does not sound correct. Most battery life tests conducted by users include things like playing videos and even games that do use the dGPU, if present. Also, websites use animations that use the dGPU.

 

[Jaekae]

I had somewhere 53-55k score before and now 47k on the 460 dGPU and this update did do some firmware changes, so maby they changed the amd firmware and clocked it down to prevent total failures of the cards or something

 

[matram]

My results from Geekbench 4

Single core
10.12.2 4.476
10.12.3 4.368

Multicore
10.12.2 13.863
10.12.3 13.560

 

[Sanpete]

That does not sound correct. Most battery life tests conducted by users include things like playing videos and even games that do use the dGPU, if present. Also, websites use animations that use the dGPU.

If people use the dGPU much in a battery test they'll come nowhere close to the numbers Apple claims. Gaming isn't part of those tests. Most videos run on the iGPU. There are things at websites that can activate the dGPU, as you say, but they don't require much from it. The benchmarks in question here are based on pushing the dGPU hard.

 

[nicolipo]

My results from Geekbench 4

Single core
10.12.2 4.476
10.12.3 4.368

Multicore
10.12.2 13.863
10.12.3 13.560

please add your geekbench openCL results, the CPU isn't affected by the update

 

[jerryk]

If people use the dGPU much in a battery test they'll come nowhere close to the numbers Apple claims. Gaming isn't part of those tests. Most videos run on the iGPU. There are things at websites that can activate the dGPU, as you say, but they don't require much from it. The benchmarks in question here are based on pushing the dGPU hard.

I agree the Geekbench benchmarks push the CPU and GPU. But many 15" buyers spent extra for GPU upgrades. So GPU performance is a concern, as is understanding if a trade-off is being made on GPU performance for battery life.
[doublepost=1485369638][/doublepost]

please add your geekbench openCL results, the CPU isn't affected by the update

Better yet, everyone should post a link to their results page with the details, like the OP.

 

[Sanpete]

So GPU performance is a concern, as is understanding if a trade-off is being made on GPU performance for battery life.

Of course. I'm pointing out that the change in performance seen here probably isn't about battery life.

 

[Creep89]

53586 @ Pro 460 (MacOS 10.12.3)
https://browser.geekbench.com/v4/compute/402362

I did not reboot, currently running on battery and I also have Football Manager 2017 still open in the background. Furthermore I do not notice any worse framerate inside Football Manager 2017.

 

[matram]

please add your geekbench openCL results, the CPU isn't affected by the update

10.12.3. 47.559 for OpenCL GPU, sorry no comparison data for 10.12.2

 

[jerryk]

53586 @ Pro 460 (MacOS 10.12.3)
https://browser.geekbench.com/v4/compute/402362

I did not reboot, currently running on battery and I also have Football Manager 2017 still open in the background. Furthermore I do not notice any worse framerate inside Football Manager 2017.

Do you have your geekbench score from 10.12.2? You have a different GPU than the OP so your numbers will be different. But the goal is to see the difference between 10.12.2 and 10.12.3.

 

[Creep89]

Unfortunately not. But now Geekbench gives me a score of about 47135. I think Geekbench is not reliable.

edit: So after I closed the game my results got weaker. Definitely weird.

edit2: Back in Football Manager and now my score is 55600. I think it is just a bug in Geekbench.

 

[matram]

GFX Metal scores if anyone is interested:

GFXBench Metal (10.12.2 / 10.12.3)
Trex 114.0 / 113.4 fps
Manhattan 53.1 / 52.8 fps
Manhattan 3.1 37.4 /37.3 fps

Offscreen
Trex 336.8 / 336.4 fps
Manhattan 177.2 / 176.7 fps
Manhattan 3.1 138.7 / 138.5 fps