This thread is now obsolete. Check out this newer thread for updated information:
So, which video card do you need? This is quite the difficult question to answer, so I did a bunch of benchmarks.
Anna Wood and I put together a set of benchmarks to see what is really required. These are meant to be real-world results. Most are from real CAD data, and the others are obviously not. I tried to be fairly comprehensive, however I did not get a chance to test every card out there. If you have data you think I should benchmark, contact me. The great thing about this being a forum is that I can continue to run the study into the future.
I do want to thank the support from @Xi computers for providing these video cards for benchmarking. I was able to benchmark the following video cards:
AMD FirePro v4900 $155
AMD FirePro W5000 $450
AMD FirePro W7000 $770
nVidia Quadro 2000 $400
nVidia Quadro 4000 $700
nVidia Quadro K5000 $1800
You can download the benchmarks I used here: http://swtuts.com/?p=700
This includes the Spin 50 benchmark macro, which was used for all the tests. Unfortunatly I can't provide all the CAD data, as some of it is proprietary. Basically, all of these tests use the "middle mouse button rotate" with an automated macro of different assemblies, and measured how many frames per second the video card could output. They were all done on one machine, so ther other hardware would be fair. They all had the same system settings in SolidWorks for image quality.
Note that everyone with a new-ish video card should turn on full screen antialiasing. This has no noticable effect on video card performance. Also, all tests were done with realview turned on.
Here are the results. Remember that video card performance is measured in frames per second. So the more FPS, the better the video looks. 12-14 FPS is the limit for human eyes to decipher something as continuous motion. After 24 fps things start to look good (TV and most movies are 24 fps), and after 30 fps they look really nice. After 60 fps things look perfectly fluid, and that is what almost all LCD monitor refresh rates are, so anything past 60 is useless.
As you can see, with massively large assemblies of the same component (a bunch of nuts), the high-end video cards did nothing. Note that transparencies don't properly utilize the video card, so it puts more load on the CPU. Since the CPU is already maxed out, it slows down the FPS. The same with edges shown, both in SW 2012 and SW 2013 (they both had similar results). So, even though this thread is about comparing video cards, if you want to speed up your assembly rotation speeds, turn off edges, and don't have transparent parts. The results of this graph show that although the bottom of the line cards are equal with the higher-end cards, somehow the AMD cards are getting data faster from the CPU, so they are performing better. Note that with edges hidden, the AMD cards look very fluid, while the nVidia cards still look quite choppy.
This data set is a large potato farm. It is almost 5GB worth of SolidWorks files. I wasn't surprised at all that this maxed out the video cards. I was surprised that the more expensive cards provided absolutely no benefit. The biggest winner was actually the cheapest card, the AMD FirePro v4900, however I think it's lead is probably an anomoly. I ran each card test 5 times and then averaged, but still saw this anomoly. I'm not quite sure why. Regardless, they are all about the same, with the AMD cards getting a slight lead.
This is the dining room set I designed for myself. All-in-all, this is a fairly simple model. but just complex enough to tax the video cards when I turned on all the "pretty options". This includes ambient occlusion, shadows, and perspective. When all those things are turned on, then we start to see results that really test what the video cards can do, and the performance matches what I would typically expect. Note, however, that with all that stuff turned off, all the cards got well above that 60 fps that us humans can see. So, really, I typically model with these features turned off, and all video cards outperformed what I can see. So this is the only way I could see someone justifying a more expensive video card. If you always want to have on draft quality ambient occlusion (where the video card renders the AO while you rotate the model), and you want to have good performance while working with small-medium sized models. Then the AMD W5000 is a strong performer, without being super-rediculously priced (just moderatly super-expensive).
This is a Die Assembly, and you can see that with the effects turned on, it only slightly slows down the video performance. This is because this assembly is right on the edge of being large enough that the CPU is the bottleneck, not the video card. So here, we see flatlines on the nVidia chips, but the AMDs have that data speed so they can perform. Of course, with the tricks turned off, we immediatly hit that CPU bottleneck. So the AMD chips win in the end, and again, somehow, the v4900 was actually the top performer with the effects turned off.
This is the Cinebench benchmark: http://www.maxon.net/products/cinebench/overview.html. It is an industry standard test, and it specifically benchmarks the video cards. As you can see, we get the exact results we expect, because it tests the video cards directly. I have included this only as a reference, as I prefer to look at the real SolidWorks data.
So my final conclusions? Well, previously I used anecdotal data to recommend the v4900, but I was really stunned to see my expectations met with such astounding data. I now have confirmed many times over my belief that all SolidWorks users should only get an AMD v4900 for $155, and spend the rest of the money on the CPU. So much of this video performance was determined by the CPU speed.
Also, I have not yet had a chance to benchmark the Intel Integraded P4000 graphics. Since that performace matches these low-end professional discreet cards, I hope to include that data soon. Who knows, maybe it will even get better results than the discreet cards. I hope to do this test sometime soon (ish), but I've already spent quite a bit of time on these benchmarks, so I need a break!