GPU benchmarks: GeForce GTX 680 vs. Radeon RX 580

[MacSince1985]

I just replaced the GPU of my Mac Pro 5,1 and ran some before and after benchmarks, which may interest some. I ran both Geekbench and Cinebench suite on Sierra, High Sierra and Mojave. In the case of NVidia, I tested both the OSX driver and the web driver (when available).

There's a fairly wide range in CPU benchmarks which I'll attribute to background processes. I rebooted between each batch of tests and turned off all other apps, but did not attempt to turn off every background process.

Conclusions:

1. Drivers: Not surprisingly, the NVidia web drivers provides consistently better OpenGL. On the other hand, the Apple drivers provide better results for Metal on the NVidia card.
2. Metal performance improves with each OS version.
3. The Radeon GPU is slightly faster in OpenGL
4. The Radeon GPU is significantly (2x) faster in OpenCL and Metal

Mac Pro 5,1
Intel Xeon CPU X5690
12 Cores, 24 Threads @ 3.46 GHz
48 GB 1333 MHz DDR3

NVIDIA GeForce GTX 680 Mac Edition (2 GB)

OS X 10.12.6 + Apple driver

• Cinebench CPU = 1527 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 14.26 x
• Geekbench CPU = 23850
• Geekbench CPU (single core) = 2893
• Cinebench OpenGL = 48.55 fps
• Geekbench OpenCL = 43648
• Geekbench CUDA = 75291
• Geekbench Metal = 67150

OS X 10.12.6 + NV web driver

• Cinebench CPU = 1520 cb
• Cinebench CPU (single core) = 105 cb
• Cinebench MP ratio = 14.46 x
• Geekbench CPU = 23607
• Geekbench CPU (single core) = 2886
• Cinebench OpenGL = 50.68 fps
• Geekbench OpenCL = 43691
• Geekbench CUDA = 76488
• Geekbench Metal = 66481

OS X 10.13.6 + Apple drivers

• Cinebench CPU = 1515 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 14.22 x
• Geekbench CPU = 24477
• Geekbench CPU (single core) = 2961
• Cinebench OpenGL = 49.85 fps
• Geekbench OpenCL = 0 (failed depth of field)
• Geekbench CUDA = n/a
• Geekbench Metal = 68836

OS X 10.13.6 + NV web driver

• Cinebench CPU = 1610 cb
• Cinebench CPU (single core) = 111 cb
• Cinebench MP ratio = 14.51 x
• Geekbench CPU = 24837
• Geekbench CPU (single core) = 2923
• Cinebench OpenGL = 56.63 fps
• Geekbench OpenCL = 52189
• Geekbench CUDA = 76286
• Geekbench Metal = 65961

OS X 10.14.1 + Apple drivers

• Cinebench CPU = 1445 cb
• Cinebench CPU (single core) = 106 cb
• Cinebench MP ratio = 13.58 x
• Geekbench CPU = 25668
• Geekbench CPU (single core) = 2909
• Cinebench OpenGL = 50.99 fps
• Geekbench OpenCL = 0 (failed depth of field)
• Geekbench CUDA = n/a
• Geekbench Metal = 69221

Sapphire Pulse Radeon rx580 (8 GB)

OS X 10.12.6 + Apple driver

• Cinebench CPU = 1449 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 13.57 x
• Geekbench CPU = 23663
• Geekbench CPU (single core) = 2800
• Cinebench OpenGL = 56.41 fps
• Geekbench OpenCL = 138117
• Geekbench CUDA = n/a
• Geekbench Metal = 139747

OS X 10.13.6 + Apple driver

• Cinebench CPU = 1532 cb
• Cinebench CPU (single core) = 109 cb
• Cinebench MP ratio = 14.10 x
• Geekbench CPU = 24721
• Geekbench CPU (single core) = 2941
• Cinebench OpenGL = 53.10 fps
• Geekbench OpenCL = 138893
• Geekbench CUDA = n/a
• Geekbench Metal = 138208

OS X 10.14.1 + Apple driver

• Cinebench CPU = 1472 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 13.81 x
• Geekbench CPU = 24677
• Geekbench CPU (single core) = 2903
• Cinebench OpenGL = 56.98 fps
• Geekbench OpenCL = 140853
• Geekbench CUDA = n/a
• Geekbench Metal = 141885

 

[AidenShaw]

I just replaced the GPU of my Mac Pro 5,1 and ran some before and after benchmarks, which may interest some. I ran both Geekbench and Cinebench suite on Sierra, High Sierra and Mojave. In the case of NVidia, I tested both the OSX driver and the web driver (when available)...

It may be interesting to some, but it shouldn't be surprising that a card from spring 2012 with 2 GiB VRAM and 3.5M transistors is slower than a card from spring 2017 with 8 GiB VRAM and 5.7M transistors.

It's good info for someone looking to upgrade a cMP - but the GTX 680 is a *really* old card.

 

[thornslack]

Always nice to see graphics data though, thanks OP. I used to do game benches with Heaven, Valley, and metro last light to try and track performance deltas. I did 680, 780, 980, and Titan but then I gave up

 

[AidenShaw]

Titan but then I gave up

Which Titan - there are:

• GTX Titan, GTX Titan Black, GTX Titan X (Kepler, 700 series)
• GTX Titan X (Maxwell, 900 series)
• GTX Titan X (Pascal), GTX Titan Xp (Pascal, 1000 series)
• GTX Titan V (Volta series)
• Soon to be released RTX Titan (Turing, 2000 series) https://www.tweaktown.com/news/64034/nvidias-unannounced-titan-rtx-graphics-card-teased/index.html

 

[koyoot]

Use proper benchmarks, not complete and utter s*** ones for GPUs.

 

[AidenShaw]

Use proper benchmarks, not complete and utter s*** ones for GPUs.

That's very helpful. (Is an emoticon for sarcasm necessary here?)

 

[thornslack]

Haha so many titans these days. I did the maxwell and pascal series Titan x’s as I was lucky enough to find them on my local Craigslist for cheap. If someone wants to provide me with the upcoming RTX Titan I will gladly do benchies

 

[MacSince1985]

It's good info for someone looking to upgrade a cMP - but the GTX 680 is a *really* old card.

The MP 5,1 is even older ;-)
Many MP owners out there still have the Radeon 5770 or replaced it with a GTX 680. So having some more info about upgrades that don't break the bank is always helpful, if not earth-shattering.

 

[h9826790]

I just replaced the GPU of my Mac Pro 5,1 and ran some before and after benchmarks, which may interest some. I ran both Geekbench and Cinebench suite on Sierra, High Sierra and Mojave. In the case of NVidia, I tested both the OSX driver and the web driver (when available).

There's a fairly wide range in CPU benchmarks which I'll attribute to background processes. I rebooted between each batch of tests and turned off all other apps, but did not attempt to turn off every background process.

Conclusions:

1. Drivers: Not surprisingly, the NVidia web drivers provides consistently better OpenGL. On the other hand, the Apple drivers provide better results for Metal on the NVidia card.
2. Metal performance improves with each OS version.
3. The Radeon GPU is slightly faster in OpenGL
4. The Radeon GPU is significantly (2x) faster in OpenCL and Metal

Mac Pro 5,1
Intel Xeon CPU X5690
12 Cores, 24 Threads @ 3.46 GHz
48 GB 1333 MHz DDR3

NVIDIA GeForce GTX 680 Mac Edition (2 GB)

OS X 10.12.6 + Apple driver

• Cinebench CPU = 1527 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 14.26 x
• Geekbench CPU = 23850
• Geekbench CPU (single core) = 2893
• Cinebench OpenGL = 48.55 fps
• Geekbench OpenCL = 43648
• Geekbench CUDA = 75291
• Geekbench Metal = 67150

OS X 10.12.6 + NV web driver

• Cinebench CPU = 1520 cb
• Cinebench CPU (single core) = 105 cb
• Cinebench MP ratio = 14.46 x
• Geekbench CPU = 23607
• Geekbench CPU (single core) = 2886
• Cinebench OpenGL = 50.68 fps
• Geekbench OpenCL = 43691
• Geekbench CUDA = 76488
• Geekbench Metal = 66481

OS X 10.13.6 + Apple drivers

• Cinebench CPU = 1515 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 14.22 x
• Geekbench CPU = 24477
• Geekbench CPU (single core) = 2961
• Cinebench OpenGL = 49.85 fps
• Geekbench OpenCL = 0 (failed depth of field)
• Geekbench CUDA = n/a
• Geekbench Metal = 68836

OS X 10.13.6 + NV web driver

• Cinebench CPU = 1610 cb
• Cinebench CPU (single core) = 111 cb
• Cinebench MP ratio = 14.51 x
• Geekbench CPU = 24837
• Geekbench CPU (single core) = 2923
• Cinebench OpenGL = 56.63 fps
• Geekbench OpenCL = 52189
• Geekbench CUDA = 76286
• Geekbench Metal = 65961

OS X 10.14.1 + Apple drivers

• Cinebench CPU = 1445 cb
• Cinebench CPU (single core) = 106 cb
• Cinebench MP ratio = 13.58 x
• Geekbench CPU = 25668
• Geekbench CPU (single core) = 2909
• Cinebench OpenGL = 50.99 fps
• Geekbench OpenCL = 0 (failed depth of field)
• Geekbench CUDA = n/a
• Geekbench Metal = 69221

Sapphire Pulse Radeon rx580 (8 GB)

OS X 10.12.6 + Apple driver

• Cinebench CPU = 1449 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 13.57 x
• Geekbench CPU = 23663
• Geekbench CPU (single core) = 2800
• Cinebench OpenGL = 56.41 fps
• Geekbench OpenCL = 138117
• Geekbench CUDA = n/a
• Geekbench Metal = 139747

OS X 10.13.6 + Apple driver

• Cinebench CPU = 1532 cb
• Cinebench CPU (single core) = 109 cb
• Cinebench MP ratio = 14.10 x
• Geekbench CPU = 24721
• Geekbench CPU (single core) = 2941
• Cinebench OpenGL = 53.10 fps
• Geekbench OpenCL = 138893
• Geekbench CUDA = n/a
• Geekbench Metal = 138208

OS X 10.14.1 + Apple driver

• Cinebench CPU = 1472 cb
• Cinebench CPU (single core) = 107 cb
• Cinebench MP ratio = 13.81 x
• Geekbench CPU = 24677
• Geekbench CPU (single core) = 2903
• Cinebench OpenGL = 56.98 fps
• Geekbench OpenCL = 140853
• Geekbench CUDA = n/a
• Geekbench Metal = 141885

Thanks for the tests, but sorry to tell you that you weren't testing the OpenGL performance but the "CPU single thread performance + OpenGL driver efficiency". This is a known limitation of Cinebench.

If you want to compare GPU's openGL performance, you better run Unigine Heaven / Unigine Valley. They can still be CPU single thread limiting, but you can always increase resolution etc to make it become GPU limiting. For GTX680 / RX580, those Extreme / Extreme HD preset should be good enough to stress the GPU to 100%.

 

[MacSince1985]

Thanks for the tests, but sorry to tell you that you weren't testing the OpenGL performance but the "CPU single thread performance + OpenGL driver efficiency". This is a known limitation of Cinebench.

That explains why the difference between the OpenGL tests was minimal. Thanks.

 

[donvito4ever]

That explains why the difference between the OpenGL tests was minimal. Thanks.

And you will do this, how h9826790 said?

If you want to compare GPU's openGL performance, you better run Unigine Heaven / Unigine Valley. They can still be CPU single thread limiting, but you can always increase resolution etc to make it become GPU limiting. For GTX680 / RX580, those Extreme / Extreme HD preset should be good enough to stress the GPU to 100%.

 

[nudoru]

I have 2x x5690's and a 4gb RX 480 with 16gb RAM and my Geekbench scores have been in the 21,000's and not even 22k. Any thoughts on how to improve it? More RAM or did I mess up with the thermal paste

Just ran it and I get
Single-Core Score 2877
Multi-Core Score 21567

 

[MacSince1985]

And you will do this, how h9826790 said?

I'll see how much free time I have ...
[doublepost=1543879602][/doublepost]

I have 2x x5690's and a 4gb RX 480 with 16gb RAM and my Geekbench scores have been in the 21,000's and not even 22k. Any thoughts on how to improve it? More RAM or did I mess up with the thermal paste

Just ran it and I get
Single-Core Score 2877
Multi-Core Score 21567

These are CPU scores, not GPU. What processors do you have in your MP? RAM speed (1066 vs. 1333 MHz) could also make a difference.

 

[h9826790]

I have 2x x5690's and a 4gb RX 480 with 16gb RAM and my Geekbench scores have been in the 21,000's and not even 22k. Any thoughts on how to improve it? More RAM or did I mess up with the thermal paste

Just ran it and I get
Single-Core Score 2877
Multi-Core Score 21567

16GB RAM is the cause. My suggestion is don't worry about that score. Which is very artificial, and not quite able to reflect real world performance difference.

 

[orph]

the RX580 is relay nice have fun, it's fun to play with benchmarks but eats so much time.
with openCL (luxmark) i saw about 3x speed boost with a GTX770 4GB -> RX580 8GB so there is a big boost

in windows it's much faster which is a plus for games and such, osx game benchmarks tend to be more limited by OSX than the GPU i think (and then the cpu limits things trying to make up for osx)

the big thing i relay noticed (and wanted) was DaVinci Resolve 15 when editing & grading 4K video from my Panasonic camera became much more fluid, it's been the only upgrade iv done that has given me a real worth while speed boost in a long time
I relay dont know how much headroom i gained in Resolve but it feels like a lot

dont forget the RX580 has a lot more headroom in both speed and vram so if you can use applications that can take advantage of both you will see the most befit in daily use


blender has a benchmark now, may be worth a go
https://opendata.blender.org/

 

[nudoru]

16GB RAM is the cause. My suggestion is don't worry about that score. Which is very artificial, and not quite able to reflect real world performance difference.

Thanks for the tip! I'm not so must worried about perf, just the 4k difference and wondering if I have bad chip. A RAM upgrade is on my Christmas list so I'll have that taken care of soon enough.

 

[orph]

Geekbench score's are not something to stress over, there artificial benchmarks.
there's old benchmarks from back in the day showing that the ram speed thing gave only 1-3% speed changes in real apps that actually stress ram.

but if it's cheep and your happy to buy it grab the ram, for speed gains PCI USB 3 card/SSD's and GPU tend to be the easy gains and ram is only relay a problem if your using more than you have installed

 

[MacSince1985]

Thanks for the tip! I'm not so must worried about perf, just the 4k difference and wondering if I have bad chip. A RAM upgrade is on my Christmas list so I'll have that taken care of soon enough.

I have 48GB of RAM as 6x8GB to take advantage of the triple channels. But that does not make a noticeable difference in real life. Background processes during the tests tend to be the bigger issue, as well as temperature that may make your processor throttle down a little.

I just sold the GTX 680 for more than I paid for the RX 580, so I'm happy and unable to run more benchmarks. ;-)

 

[nudoru]

I have 48GB of RAM as 6x8GB to take advantage of the triple channels. But that does not make a noticeable difference in real life. Background processes during the tests tend to be the bigger issue, as well as temperature that may make your processor throttle down a little.

I just upgraded to 48gb (8gbx6) and my multicore shot up to 26,253.

 

[feinberj]

I just found this thread. It's a great thread, for me, as I have a k5000 for Mac card for my 5,1 Mac Pro and I'm wondering if it would be worth it to upgrade. This card is essentially a GTX 680, but with Apple-blessed firmware that gives a bootscreen and also works with Mojave with no additional flashing needed. I recently upgraded my 5,1 to Mojave (from Sierra) and have noticed that VMware Fusion seems a bit snappier when I run my virtual machines in retina mode. Under Sierra, when I turned on retina for my Win 10 VM the UI got sluggish enough that I switched retina mode back off. Now in Mojave, the Win 10 UI runs well enough with retina mode on that it's tolerable. It's been quite a while since I tried running my Win 10 VM with retina mode on, probably more than a year, so maybe it's something else that made the difference. Maybe because I'm now running a newer version of Vmware, or even a newer version of Windows.

I'm mostly doing software development in Windows, so no gaming. Just Outlook, Powerpoint, Excel, and my softawre development app. No gaming. Also I never bothered with the nVidia drivers. I'm just using whatever driver comes native with the Mac OS - first Mavericks, then Sierra, and now Mojave.

Anyhow, I was wondering if my Win10 virtual machine would "feel" noticeably snappier in retina mode if I upgraded from my nVidia k5000 to an AMD 580, such as the Sapphire Nitro 580 that's recommended for the 5,1 Mac Pro. I'm starting a project that involves generating HiDPI images for web pages on an internal corporate server, so being able to run Win 10 in retina mode has become quite useful to me. I really notice the difference between a snappy and sluggish Win 10 UI, so I'd be willing to upgrade the video card if it made my Win 10 VM "feel" faster.

 

[killawat]

Anyhow, I was wondering if my Win10 virtual machine would "feel" noticeably snappier in retina mode if I upgraded from my nVidia k5000 to an AMD 580, such as the Sapphire Nitro 580 that's recommended for the 5,1 Mac Pro. I'm starting a project that involves generating HiDPI images for web pages on an internal corporate server, so being able to run Win 10 in retina mode has become quite useful to me. I really notice the difference between a snappy and sluggish Win 10 UI, so I'd be willing to upgrade the video card if it made my Win 10 VM "feel" faster.

Give your VM as much RAM and CPU as possible (without affecting macOS) to see if you Can get acceptable levels without upgrade the GPU. But I run several VMs across multiple 4k displays on macMini + 580 eGPU with no effect.

 

[orph]

you may be better of looking at some fast drives over a GPU.