Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Pascal OpenCL performance/issues

AndreeOnline

AndreeOnline

macrumors 6502a
Original poster
Aug 15, 2014
714
507
Zürich
I just popped a 1080 Ti into my cMP. Drivers install smoothly, both the main Nvidia package and the subsequent CUDA update.

Geekbench Metal and CUDA run flawlessly. DaVinci Resolve with CUDA runs great as well. As does Heaven Benchmark.

But I feel there is something wrong with OpelCL. Geekbench 4.1 OpenCL test drops out when it hits the second 'histogram' test and fails the benchmark. Luxmark also fails at least the Lobby scene (I get scores for the Ball and Mic). DaVinci Resolve from MacAppStore grinds to a halt in OpenCL mode, and that version doesn't support CUDA (Doh!). Performance in FCPX is also weak.

I was under the impression that OpenCL performance with other Nvidia cards like Maxwell had recovered nicely and was in the "normal/expected" range?

My poor performance together with failing benchmarks points to a more serious issue. I first thought it had something to do with my old .kext mods for my RX 480, but I've made a new MacOS upgrade install and nothing changes there.

I know there are other threads discussion general performance of these cards/drivers, but I'd like to isolate this specific issue and see if it's the same for everyone, or if I have some other issue. My card has enough power and is completely stable in other benchmarks.

If you have similar issues, please report. Thanks!
 
AndreeOnline said:
I just popped a 1080 Ti into my cMP. Drivers install smoothly, both the main Nvidia package and the subsequent CUDA update.

Geekbench Metal and CUDA run flawlessly. DaVinci Resolve with CUDA runs great as well. As does Heaven Benchmark.

But I feel there is something wrong with OpelCL. Geekbench 4.1 OpenCL test drops out when it hits the second 'histogram' test and fails the benchmark. Luxmark also fails at least the Lobby scene (I get scores for the Ball and Mic). DaVinci Resolve from MacAppStore grinds to a halt in OpenCL mode, and that version doesn't support CUDA (Doh!). Performance in FCPX is also weak.

I was under the impression that OpenCL performance with other Nvidia cards like Maxwell had recovered nicely and was in the "normal/expected" range?

My poor performance together with failing benchmarks points to a more serious issue. I first thought it had something to do with my old .kext mods for my RX 480, but I've made a new MacOS upgrade install and nothing changes there.

I know there are other threads discussion general performance of these cards/drivers, but I'd like to isolate this specific issue and see if it's the same for everyone, or if I have some other issue. My card has enough power and is completely stable in other benchmarks.

If you have similar issues, please report. Thanks!
Click to expand...

Search my screen name+Nvidia+open cl

Been saying this for almost two years and said it was getting worse with each new driver. Nvidia enable new generation of cards but they don't optimise for them or ever come out of beta. This is essentially still a Kepler driver running on newer architecture via the UDA core.
 
AndreeOnline said:
That might very well be the case, but with each new driver the expectations is that they will have caught up. Especially now that Nvidia is beating the marketing drum.

If they know about it, and can fix it, I'm guessing they will.
Click to expand...

The Nvidia, Hackintosh and Adobe forums have been full of complaints since 2015 and nothing changes. They should also be aware of discussions here too.

That's why I just gave up on cMP. Sold it, put my Nvidia cards in a PC and use a MacBook Pro until Apple addressed this in 2018 or whenever.

There are ridiculous video bugs in nMP so MBP was my only Mac choice.
 
Fair enough. Depending on where we all enter and exit the 'timeline/arc' for a certain phenomenon , we'll make our own experiences that lead to the decisions we make.

I'm confident you have your own very clear reasons for what you chose, and that is all that matters.
[doublepost=1492861601][/doublepost]In lack of other tests, I downloaded Compubench OpenCL, GFXBench Metal and OpenGL.

The tests run normally and my configuration places near the top, indicating that there is nothing 'wrong' with my computer or card. It's just a question of compatibility.

It really seems to be a situation where it won't help if you throw money at the problem.
 
DearthnVader said:
Just to be clear, WebDriver 378.05.05f02 + Cuda Driver 8.0.81?
Click to expand...

CUDA works great. I found a slightly alternative workflow with C4D and Maxwell where my renders are basically 10x faster on the 1080 Ti compared to 12c 3.46GHz CPU. That is quite a punch that is hard to let go of....
 
  • Like
Reactions: Synchro3
DearthnVader said:
The problem could be that nVidia seems to be using OpenCL 1.1, while Apple/LuxMark seem to be using OpenCL 1.2.

Try https://github.com/zbendefy/opencl-z/releases/download/1.1b/opencl_z_1_1b.jar, and see what version of OpenCL is reported for your 1080ti.
Click to expand...

Just used it to check my GTX 780. (My GTX 1080 is on order) Looks like it's OpenCL 1.2

TinyGrab Screen Shot 4-22-17, 10.40.05 AM.png

Lou
 
Just thought I'd alert everyone here. I just got off MacVidCards blog, and there was a message posted today by MVC that he is finishing up testing of the EFI flash for the GTX 1080 family of GPU's. He expects to start shipping product by the first of May:

http://www.macvidcards.com/blog

Lou
 
flowrider said:
Just thought I'd alert everyone here. I just got off MacVidCards blog, and there was a message posted today by MVC that he is finishing up testing of the EFI flash for the GTX 1080 family of GPU's. He expects to start shipping product by the first of May:

http://www.macvidcards.com/blog

Lou
Click to expand...

Remarkable. This thread outlines bugs that impact professionals and you are deflecting the topic into a sales opportunity.
 
No need to bump heads guys. At some point we all came here for similar reasons. If you for one reason of another have had bad experiences, it's only natural to feel frustrated. But some of the discussions I see lately are unlikely to get "resolved".

Let's remember that we on the Mac Pro forum is an almost negligible subset of Mac users that face more adversity than most. That said, we who DO come here are looking for solutions, good ideas or just plain old inspiration.

Anyone who is over the Mac Pro and done with it don't have to piss in the punch for those of us who try to make the most of it. If for no other reason than common curtesy and respect.

Heads high and level, gentlemen.
 
  • Like
Reactions: phil_enn, mostafiz28, pastrychef and 4 others
SoyCapitanSoyCapitan said:
Search my screen name+Nvidia+open cl

Been saying this for almost two years and said it was getting worse with each new driver. Nvidia enable new generation of cards but they don't optimise for them or ever come out of beta. This is essentially still a Kepler driver running on newer architecture via the UDA core.
Click to expand...

I can't private message you, but would like to talk more about it for a news story I'm writing. Very interesting stuff.
 
Anyway, back to the topic.

It seems nVida OpenCL performance is just fine on Windows:

https://compubench.com/result.jsp

, so I really don't understand why it's so poor on the MacOS. Is it drivers, or is it Apple's implementation of OpenCL?

I really don't know, but I can't imagine the people that write the MacOS drivers over at nVidia are sitting around saying, " Boy, let's just put out the crappiest drivers we can for OpenCL on the Mac, because we hate OpenCL on the Mac."

Kind of leads me to believe it Apple's implementation of OpenCL, that for reasons of wanting to sell new Macs, that all come with AMD hardware, they have "optimized" their OpenCL implementation for AMD, but that's just pure speculation on my part.

The only real way we'll know for sure is if nVidia or Apple start talking, and that isn't likely to happen, as nVidia wants to sell OEM to Apple again, and doesn't want to risk the prospect of losing hundreds of millions in sales, by opening their mouth to bad mouth Apple. Apple would likely never admit they were manipulating their OpenCL implementation to favor AMD hardware, in order to drive sales of new Macs, not only because they risk alienating customers, but also because Apple has never really like being that open or commutative about issue with anything, much less, someone else's hardware.
 
DearthnVader said:
Anyway, back to the topic.

It seems nVida OpenCL performance is just fine on Windows:

https://compubench.com/result.jsp

, so I really don't understand why it's so poor on the MacOS. Is it drivers, or is it Apple's implementation of OpenCL?

I really don't know, but I can't imagine the people that write the MacOS drivers over at nVidia are sitting around saying, " Boy, let's just put out the crappiest drivers we can for OpenCL on the Mac, because we hate OpenCL on the Mac."

Kind of leads me to believe it Apple's implementation of OpenCL, that for reasons of wanting to sell new Macs, that all come with AMD hardware, they have "optimized" their OpenCL implementation for AMD, but that's just pure speculation on my part.

The only real way we'll know for sure is if nVidia or Apple start talking, and that isn't likely to happen, as nVidia wants to sell OEM to Apple again, and doesn't want to risk the prospect of losing hundreds of millions in sales, by opening their mouth to bad mouth Apple. Apple would likely never admit they were manipulating their OpenCL implementation to favor AMD hardware, in order to drive sales of new Macs, not only because they risk alienating customers, but also because Apple has never really like being that open or commutative about issue with anything, much less, someone else's hardware.
Click to expand...

If OpenCL is working fine in MacOS with AMD. Nvidia has always been concentrating on CUDA and anything derived from it. They let OpenCL to the wayside and now that OpenCL is the main use in MacOS and their default applications, its likely Nvidia is still catching up. Could be a quick fix just to get it up and running with newest cards until better drivers come out.
 
Here are my results compared to generic Windows. The rest of the computer shouldn't matter too much since these tests run pretty isolated on the GPU.

Whatever the reason (and I don't believe it's a conspiracy), I hope it can get fixed. And if Apple needs to do something on their end, they really should.

Screen Shot 2017-04-23 at 21.38.15.png
 
  • Like
Reactions: thornslack
AndreeOnline said:
Here are my results compared to generic Windows. The rest of the computer shouldn't matter too much since these tests run pretty isolated on the GPU.

Whatever the reason (and I don't believe it's a conspiracy), I hope it can get fixed. And if Apple needs to do something on their end, they really should.

View attachment 697286
Click to expand...
At least the tests complete for you, what benchmark is that?

I tried the CompuBench from the AppStore, and under 10.12.4 the test would render fine, but the app would quit after each test.

Under 10.11.1 with the nVidia web drivers it runs fine.
 
Ok, I don't know how you got that nice side by side of Compubench 1.5?

Here are the results of some of my finding so far, not that they really apply to Pascal, but I'm just trying to narrow down OpenCL issues with nVidia under the MacOS.

Quadro 600

First Luxmark V2.0: I rendered the scene Sala

Linux Ubuntu 14.01: nVidia Driver 378.xxx Score 36

OS X 10.11.1 nVidia WebDriver 346.03.03f02 Score 126 CPU Score 176

OS X 10.12.4 Default OS X nVidia driver Score 127 CPU Score 170

Windows 8.1 Quadro Driver Score 190

Luxmark v2.0 Pegged the GPU Utilization at between 97-99% in Windows, and 99% in OS X, but under Linux it never got higher than 14%, even when I turned the Powermizer settings to Max Performance.

Compubench v1.5

OS X 10.11.1 nVidia WebDriver 346.03.03f02

Face Detection 7.7533
TV-L1 1.7863
Ocean Surface 205.55
Partical Stimulation 56.151
T-Rex 0.56731
Video Composition 9.5232
Bitcoin Mining 11.662

Windows 8.1 Quadro Driver
Face Detection 5.149
TV-L1 2.5749
Ocean Surface 236.73
Partical Stimulation 63.017
T-Rex 0.61898
Video Composition 11.272
Bitcoin Mining 19.51

Seems most of these are within the 15% overhead of the nVidia Drivers having to call on Apple's OpenCL framework on the MacOS, rather than the more direct way in Windows. I don't believe Wondows or Linux has it's own OpenCL impamention, nVidia controls the entire stack.

**Note that with the Default OS X nVidia driver, under both 10.11.1 and 10.12.4, I was unable to benchmark Compubench, the scenes would render, but Compubench would quit between each benchmark, and never return a score. I am unable to run the nVidia WebDriver for 10.12.x, as it has a jpeg rendering bug on my Quadro 600.

More data is needed, for other Fermi, kepler, maxwell, and pascal users. Maybe if we can collect enough data, we can get nVidia to fix the trouble with Maxwell and Pascal.

To get a nice histogram of GPU Utilization when you run these benchmarks, under OS X, download Graphics tools for Xcode and open the OpenGL Driver Monitor and add GPU Core Utilization and set the Min Value to 10M.
http://adcdownload.apple.com/Develo...or_Xcode_7.2/Graphics_Tools_for_Xcode_7.2.dmg
http://adcdownload.apple.com/Develo...or_Xcode_7.2/Graphics_Tools_for_Xcode_7.2.dmg
 
Last edited:
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back