Recommend the best GPU for Mac Pro

[Jon MIchael Puntervold]

I ran almost every benchmark and diagnostic tool I could find today, and from the results I gathered, the performance on an unmodified GTX580 seemed on par with it's windows brethren and PCI-bus speed is above Mac Quadro 4000. OSX 10.8.2 and latest nVidia drivers seems to be able to run the GTX cards normally (except the boot-up screen).

 

[Jeff Kilgroe]

Look in your system profiler to see what the PCIe link speed is for the card. If it reports back as 5.0GT/s, then you're running at the proper speed. If it reports back as 2.5GT/s, then you're not and are locked into low-power mode. As for GPU benchmarks, which ones are you running? I'm not aware of any good ones on the Mac, short of stress-testing the GPU in Resolve. Don't waste your time with Cinebench, they haven't updated their tests in ages and they don't stress any of the newer hardware or give any meaningful anything.

 

[Ryan De Franco]

Ah but Jeff, we sad hacks with our un-flashed 570s are unable to review PCIe data. "There was an error while gathering PCI card information." Another sad side effect of no Apple ROM, even with updated CUDA drivers.

Sadly no PCIe link speed under the Graphics tab either.

I wonder if there is another utility that can check PCIe link speed, or a common benchmark someone with a flashed 570 might be kind enough to run?

 

[Jon MIchael Puntervold]

Ryan, correct, no system profiler for us...

CUDA - Z's should give you a proper view of the PCI bandwidth.
Nice visual bench mark: http://unigine.com/products/valley/
Nice OpenCL render with comprehensive performance database: http://www.luxrender.net/wiki/LuxMark

 

[Jon MIchael Puntervold]

I wonder if there is another utility that can check PCIe link speed, or a common benchmark someone with a flashed 570 might be kind enough to run?

Would be great if someone could double check for us, but I'm pretty sure I've read somewhere that with the current drivers GTX5xx cards are fully supported, both in terms of power management and proper bandwidth.

 

[Jeff Kilgroe]

Ah but Jeff, we sad hacks with our un-flashed 570s are unable to review PCIe data. "There was an error while gathering PCI card information." Another sad side effect of no Apple ROM, even with updated CUDA drivers.

Sadly no PCIe link speed under the Graphics tab either.

I wonder if there is another utility that can check PCIe link speed, or a common benchmark someone with a flashed 570 might be kind enough to run?

Heh. Yeah, I guess this does happen with many cards -- I forgot. It used to list the link speed under the graphics tab in prior OSX versions. It's also difficult to determine what's going on at all times and you have to pull a PCI table at the time the card is being used in a game or other app that has it ramped up. When at boot or at idle, the link speeds will usually report 2.5GT/s because they throttle down, even when being used properly.

There's a utility called lspci, popular with the hackintosh crowd, that queries and builds a PCIe device table. However, when you have a blank EFI entry for GPU, which causes the PCIe polling error -- which you are seeing in Profiler, it may not work. I would see what CUDA-Z reports in the host to device score for memory copy and then maybe compare to the score on a Windows system and see if there's any discrepancy, although none of those speeds are going to stress any PCIe interface, even running at 2.5GT/s. The bus speed isn't the real issue here, it's making sure the card is coming out of low power mode so that it runs at optimal clock speed for the task at hand. CUDA-Z is helpful as it will report clock speeds and you can monitor. If the GPU clock ramps up and down along with the workload while you hit the GPU hard in an app like Resolve, then you can bet it's probably working. Also be sure to check for proper fan speed control. There's another utility that can check PCIe link and other attributes, but I'm not thinking of it off the top of my head...

As for benchmarks, the best one is still to stress-test in Resolve. Most other CUDA, OpenCL/GL, etc.. benchmarks are very small when it comes to size and they don't truly test the card under any strain. CUDA-Z rates execution speed for CUDA operations, but doesn't cycle enough data to stress any PCIe interface or the onboard RAM and bank caches for the GPU nodes. It's handy for saying that some new GPU is faster than it's predecessor or other similar GPUs in the line-up, but it doesn't give us the whole picture.