2008 Mac Pro vs. 2016 Dell Precision Performance
I did a quick experiment comparing performance with C1 9 on my work versus home computer. This isn’t a question per-se, but it could help others troubleshooting performance issues. It’s rather enlightening considering the variables. My home computer is an early 2008 Mac Pro and my work computer is a new custom build Dell Precision T1700, specs below.
I created a small standalone catalog with 23 Nikon D750 RAW images with a typical number of adjustments for me. All have adjustments for clarity, highlights, sharpening, noise reduction, lens correction (distortion and chromatic aberration), and vignetting. Some images also have exposure adjustments and additional shadow adjustments, as well as straightening and keystone. I bundled the RAW files into the catalog in this case to make sure it was easily portable between computers, though I usually use referenced files. Since both computers are booting off an SSD and this is a brand new catalog with no other images, disk access shouldn’t be a factor, just CPU and GPU, as well as any Mac/Windows variables.
I tested exporting the images to full-size JPEG files at 80% quality with metadata and color profiles intact.
Mac Pro 3,1 Early 2008
8-core (2x 4-core) 2.8 GHz Xeon Harpertown (No Hyperthreading, No Turbo-Boost)
ATI Radeon HD 5770 1GB
18 GB ECC RAM
256 GB OWC Mercury Accelsior PCIe SSD Boot Drive
OS X 10.10.5 Yosemite
27†LED Cinema Display 2560x1440
Dell Precision T1700
4-core 3.5 GHz Xeon Haswell (Hyperthreaded, 3.9 GHz Turbo-Boost)
Nvidia Quadro K620 2GB
16 GB ECC RAM
360 GB Intel SSD Boot Drive
Windows 7 Pro SP1
(2) Dell P2213 Monitors 1680x1050 each
Here are the times for the two computers:
Mac Pro
Open CL Off 2:03 (85% cpu on 8 threads)
Open CL On 4:57 (minimal cpu)
Dell
Open CL Off 2:00 (93% cpu on 8 threads)
Open CL On 1:12 (minimal cpu)
I knew going in that Open CL on the Mac was not going to be good. It makes little to no difference performance-wise for image viewing, but for exporting it’s a total dog and it even bogs down the rest of the computer. The Radeon HD 5770 just isn’t up to the task, which is to be expected.
What’s really interesting is that in raw CPU power the two computers are basically neck-and-neck. It goes to show that in this particular case, hyperthreading is only boosting performance on the Dell by about 40% versus a theoretical 100%. So it’s similar to the performance of a 6-core processor without hyperthreading. Note that I’ve yet to see the Dell throttle down below 3.9 GHz when under load, it seems to have a good enough heat sink to enable it to stay at peak frequency.
You can also see that the Nvidia Quadro K620 really chews through the files nicely. Again I don’t notice much difference on the Dell in viewing performance with Open CL on or off but scroll wheel zooming is much better across the board on the Dell than on the Mac Pro.
I can see why PhaseOne pushes for the best video card possible, since if C1 can take advantage of their modern processing capabilities, it can really sing. It makes me think that a 4-core Xeon or i7 with a tricked out GPU (or two) would be a better performer with C1 than a 6-core or 8-core Xeon with lesser GPU power.
I created a small standalone catalog with 23 Nikon D750 RAW images with a typical number of adjustments for me. All have adjustments for clarity, highlights, sharpening, noise reduction, lens correction (distortion and chromatic aberration), and vignetting. Some images also have exposure adjustments and additional shadow adjustments, as well as straightening and keystone. I bundled the RAW files into the catalog in this case to make sure it was easily portable between computers, though I usually use referenced files. Since both computers are booting off an SSD and this is a brand new catalog with no other images, disk access shouldn’t be a factor, just CPU and GPU, as well as any Mac/Windows variables.
I tested exporting the images to full-size JPEG files at 80% quality with metadata and color profiles intact.
Mac Pro 3,1 Early 2008
8-core (2x 4-core) 2.8 GHz Xeon Harpertown (No Hyperthreading, No Turbo-Boost)
ATI Radeon HD 5770 1GB
18 GB ECC RAM
256 GB OWC Mercury Accelsior PCIe SSD Boot Drive
OS X 10.10.5 Yosemite
27†LED Cinema Display 2560x1440
Dell Precision T1700
4-core 3.5 GHz Xeon Haswell (Hyperthreaded, 3.9 GHz Turbo-Boost)
Nvidia Quadro K620 2GB
16 GB ECC RAM
360 GB Intel SSD Boot Drive
Windows 7 Pro SP1
(2) Dell P2213 Monitors 1680x1050 each
Here are the times for the two computers:
Mac Pro
Open CL Off 2:03 (85% cpu on 8 threads)
Open CL On 4:57 (minimal cpu)
Dell
Open CL Off 2:00 (93% cpu on 8 threads)
Open CL On 1:12 (minimal cpu)
I knew going in that Open CL on the Mac was not going to be good. It makes little to no difference performance-wise for image viewing, but for exporting it’s a total dog and it even bogs down the rest of the computer. The Radeon HD 5770 just isn’t up to the task, which is to be expected.
What’s really interesting is that in raw CPU power the two computers are basically neck-and-neck. It goes to show that in this particular case, hyperthreading is only boosting performance on the Dell by about 40% versus a theoretical 100%. So it’s similar to the performance of a 6-core processor without hyperthreading. Note that I’ve yet to see the Dell throttle down below 3.9 GHz when under load, it seems to have a good enough heat sink to enable it to stay at peak frequency.
You can also see that the Nvidia Quadro K620 really chews through the files nicely. Again I don’t notice much difference on the Dell in viewing performance with Open CL on or off but scroll wheel zooming is much better across the board on the Dell than on the Mac Pro.
I can see why PhaseOne pushes for the best video card possible, since if C1 can take advantage of their modern processing capabilities, it can really sing. It makes me think that a 4-core Xeon or i7 with a tricked out GPU (or two) would be a better performer with C1 than a 6-core or 8-core Xeon with lesser GPU power.
0
-
Does the time C1P needs for exporting JPG images on the MAC pro with OpenCL enabled increase over time, i.e. does the 20th image take much longer than the 1st or 2nd? If you see the same bug on your system as I see on my iMAC (NVIDIA GTX 680 MX, 2 GB) just looking at the time of export in finder will show the difference. 0 -
Very interesting. I only didn't understand this: [quote="jjakucyk" wrote:
It goes to show that in this particular case, hyperthreading is only boosting performance on the Dell by about 40% versus a theoretical 100%.
40%: what do you compare with?
And why do you expect 100% theoretical boost?
The Mac runs 8 threads and the Dell runs 8 threads too, so the difference is the CPU clock 3.9/2.8 - 1 = 40%. But in fact the Mac is equally fast without OpenCL in your test.
Cheers,
BeO0 -
[quote="FL" wrote:
Does the time C1P needs for exporting JPG images on the MAC pro with OpenCL enabled increase over time, i.e. does the 20th image take much longer than the 1st or 2nd? If you see the same bug on your system as I see on my iMAC (NVIDIA GTX 680 MX, 2 GB) just looking at the time of export in finder will show the difference.
I don't think so. When the export started C1 reported about 5 minutes remaining, and it seemed to progress equally slowly throughout, but I didn't watch that closely.[quote="BeO" wrote:
Very interesting. I only didn't understand this:[quote="jjakucyk" wrote:
It goes to show that in this particular case, hyperthreading is only boosting performance on the Dell by about 40% versus a theoretical 100%.
40%: what do you compare with?
And why do you expect 100% theoretical boost?
The Mac runs 8 threads and the Dell runs 8 threads too, so the difference is the CPU clock 3.9/2.8 - 1 = 40%. But in fact the Mac is equally fast without OpenCL in your test.
I didn't expect a 100% boost, that's why I said it's theoretical. If hyperthreading could actually double the multi-threaded performance of a processor, and I understand that in certain video encoding scenarios it gets pretty close to that, then the Dell should have been significantly faster than the Mac.
I tried disabling hyperthreading on the Dell, and the results are as follows:
OpenCL On 1:10
OpenCL Off 1:43
So it seems that hyperthreading actually hurts performance in this case. So what's going on then? Well, after nearly 10 years of development, the Haswell chips are significantly more efficient/powerful per clock cycle than Harpertown. If there were such a thing as a 3.9 GHz Harpertown Xeon, then a 4-core system would still take almost three minutes to process these photos. General CPU benchmarks show an approximately 70% speed improvement at the same clock frequency between Harpertown and Haswell, and this test would seem to bear that out. Fascinating.0 -
It would be interesting to see what the difference would be if you took the Quadro card from the Dell and stuck it into the Mac Pro. Very close to getting a Quadro for my mac 3,1 and interested to see if there are any issues 0 -
Would the Quadro K620 work in this Mac Pro? I know you generally need to flash them to get boot screens and all that, but I'd probably need to find a power cable for it (the Radeon has a separate power cable too) and also a DisplayPort to Mini DisplayPort adapter. I generally haven't looked at more recent cards because they require Mountain Lion to run, and I still occasionally need to boot into Snow Leopard. Though I suppose it might be possible to install both and just swap monitor cables as necessary if it turned out to be a big benefit, it's not as if it's an expensive card. Slot 3 with 4x PCIe lanes should still have enough bandwidth to handle the SSD without a performance hit compared to the 16x lanes in slots 1 and 2. The trick would be if the Quadro is small enough to not block the fan on the Radeon. You can see that even the small SSD card covers about 1/4 of it.
http://jjakucyk.com/captureone/IMG_1666.jpg
http://jjakucyk.com/captureone/IMG_1670.jpg0 -
I finally had a chance to try the Quadro K620 in my Mac Pro, and while there was promise, it ultimately flopped.
For starters, the Quadro does not need a separate power cable, so all I had to get was a DisplayPort to Mini DisplayPort adapter. I got a Startech one from NewEgg, there's surprisingly few out there. I'd downloaded and installed the appropriate Nvidia drivers already (harder than it would seem, because they have drivers that are specific not only to the operating system version (10.10.4, 10.10.5, etc.) but to the specific build installed on your machine. There's like 5 or 6 different versions just for Yosemite 10.10.5 for instance, and you need to get the right one.
Anyway, the computer booted fine, without the boot screen of course (I do not think a Mac EFI firmware is available for this card regardless). There was some speckling that I could see in dark areas of the screen. This could be a symptom of the cheap DisplayPort adapter, but I don’t have a way to test that. Still, not a good sign.
As far as CaptureOne, FL was on to something, I saw the same problem. The first 12 images took just 0:46, so extrapolating that out the whole export should’ve only taken 1:32 (nearly as fast as the Dell). Unfortunately, no such luck. It started to slow down a bit after the half way point, but it was totally in the weeds by about image 19 out of 23. The last two were painful and the whole export took 2:48, so it was still significantly longer than CPU-only. If I immediately restarted the export, then it was slow from the get-go, estimating no less than seven minutes to finish.
I know that iMacs have had issues with thermal throttling of the GPUs, but would that be likely in this case? I assume based on the small fan (and overall small size of) the K620 that it’s probably a fixed speed, but there was no evidence of throttling on the Dell so what gives? Aside from that, it didn’t improve the overall responsiveness of anything I could test. Graphics performance of Cities Skylines did not seem affected markedly either, though this really isn’t a gaming GPU.
There are certainly be some good GPU upgrades out there, but at least in this case, I don’t think the Quadro K620 is a worthwhile endeavor.0
Post is closed for comments.
Comments
6 comments