NVIDIA's Quadro K5000 in Review
Kurt Foster [Modulok] - Staff Columnist
The Quadro K5000 is a new high-end graphics card from NVIDIA. The Quadro product lineup is targeted toward professional Workstation graphics and extremely high resolution multi-device displays. If your daily routine involves Maya, SolidWorks, CATIA, AutoCad, medical imaging, gas, etc, with millions of polygons, gigabytes of textures - a Quadro card makes life easier (and in some cases possible at all - more on that later). For Adobe users, including Photoshop CS6 and many other programs in the CS6 lineup, they can also benefit from the improved compute performance of the NVIDIA Quadro K5000.
At first, the Quadro K5000 looks similar to its sibling the Quadro 5000 (note the lack of the letter 'K' in the older card's name). There are some differences in the plastic heat sink shroud. The Quadro K5000 is also a bit longer.
Other visible differences include the Quadro K5000 has 2x DVI connectors (1x DVI-D, 1x DVI-I) in addition to its 2x DisplayPort connectors, while the Quadro 5000 had only one DVI connector and 2x DisplayPort connectors. The important thing, however, is with the newer Quadro K5000 you can use all four outputs simultaneously. For example, on my test machine I had a desktop spanning 3 monitors (it would have been four, had I a spare fourth). Its predecessor only permitted two outputs at a time, despite having three headers.
The performance of the two cards is quite different. Here's why: The older Quadro 5000 is based on the Fermi architecture. While a decent performer, the Fermi based Quadro 5000 ran hotter and less efficiently. The Fermi card consumed more power (152 watts vs. 122 watts) and had a lower overall performance. As a consequence, the Fermi-based card was also a bit louder at times due to the added fan noise to keep it within their ideal operating temperatures. In a noisy office environment you probably won't notice much difference.
I was pleased to find the newer Quadro K5000 runs much cooler, a little quieter and also quite a bit faster. It employs the newer, more energy efficient, Kepler architecture. A quick look at the averaged number of gflops per-watt between the two cards is impressive (admittedly this is a somewhat naive comparison, as the two cards have different idle power consumption that was not accounted for. This was computed as flops (single-precision) / peak draw):
I found the Quadro K5000 to be the superior card in just about every quantitative test (with the exception of one, which we'll get to in a moment). Under heavy graphic load the new Quadro K5000 is no louder than your average GeForce card. Pretty quiet overall. Acoustics aside, there are some major differences between the two cards. First, a quick glance at some performance-critical numbers as reported by NVIDIA:
We can't directly compare the number of CUDA cores from two different architectures as a measure of performance. We would be comparing apples to oranges. Still, the real-world performance difference, as well as the synthetic benchmark performance difference was significant.
I benchmarked the two cards in SPECviewperf® 11. For those more familiar with the GeForce lineup, SPECviewperf® is the workstation analog of 3DMark. It's pretty much the industry standard benchmark for workstation graphics cards. One must keep in mind, however, SPECviewperf® 11 is a synthetic benchmark. The reported performance numbers will vary from one machine to another, even with identical graphics cards. This is not unexpected. The best test of performance is with real-world data that the end-user will actually be working with. First, some numbers as reported by NVIDIA:
Unfortunately, my results were not quite as awesome. Graphics cards are typically benchmarked on the latest computational monsters, so as to ensure the only bottleneck becomes the card itself. My test machine didn't do as hot as the NVIDIA test machine, however the results are still impressive. The relative performance difference between the two cards also remains. Here's mySPECviewperf® 11 results:
The results are quite similar to those of NVIDA even if, admittedly, slightly lower. Again, this is not unexpected. The specs of the test machine I used for the results displayed above, are shown below:
Obviously the Quadro K5000 is faster than the Quadro 5000. In fact, it's faster in almost every testable way. Surprisingly, however, the sw-02 test actually ran a little slower on the newer card, at least on my machine (a value of 44.16 for the Quadro 5000 vs. 43.91 for newer Quadro K5000, a slowdown of 0.25).
My SPECviewperf® 11 settings for all tests conducted using the same driver version 188.8.131.520, with all other system settings other than the card itself being identical. The results are the average of 3 iterations of the SPECviewperf® 11 test suite:
Bonus - Quadro vs. GeForce
Ever wonder why you need a Quadro for optimal workstation performance? Just for fun I ran SPECviewperf® 11 with a GeForce GTX 480. This card was also based on the Fermi architecture, just like the Quadro 5000. This should give some perspective to the workstation performance differences between the GeForce and Quadro lineup:
It's easy to think the higher performance in workstation graphics would translate into the Quadro card being also superior for gaming graphics. This is not the case. In fact, the performance on 3DMark was actually far worse.
Moral of the story: Workstation graphics and gaming graphics are entirely different. They necessitate a different approach for optimal performance. Need a high end gaming machine? Go GeForce. For gaming graphics they're actually superior. Need a workstation to precisely wield extremely high resolution, complex data sets? Go Quadro. For workstation graphics, they are vastly superior.
The subjective real-world performance of the Quadro K5000 in the applications I regularly use and know (Autodesk Maya 2013, Adobe Photoshop CS6, Autodesk Mudbox 2013), was impressive. Just as the SPECviewperf® 11 results indicate, the performance of the newer card was indeed superior. With heavy models it felt quite a bit snappier in the Maya viewport. I found I could effectively work with models up to around 20 million polygons. After that, I ran out of main system memory.
My experience with Autodesk Mudbox was similar, but with Mudbox I ran out of memory at around 135,528,448 polygons instead. The sculpting was terribly slow, the undo painfully slow, but orbiting the model remained at around 67 fps. It sounds horrible, but at 135 million polygons it's impressive that it works at all. Had I more system memory, I think I could have pushed things a lot further.
Both the Quadro 5000 and its successor the Quadro K5000 are very nice cards, and at a retail price of around $1,700.00 and $1,780.00 respectively, they ought to be. Is the Quadro K5000 worth the extra $80-$100 over its predecessor? Only you can answer that. If you're spending just shy of $2,000 on a graphics card - I'd argue yes. Its added performance with Maya for heavy data sets was worth an added $80. This was especially the case with dense and numerous textures. However, don't count the Quadro 5000 out. As many firms update to the Quadro K5000 I wouldn't be surprised to see the second hand market for the Quadro 5000 see some very attractive offerings.
You can find out more about these cards at the NVIDIA website, as well as more about the performance test suite SPECviewperf® 11 at the following links:
All supporting images are copyright, and cannot be
Kurt Foster (Modulok) falls somewhere between programmer and visual effects artist. When not sifting through technical manuals, he takes on freelance roles in both programming and visual effects, attempting to create a marriage of technical knowledge with artistic talent. He can be seen helping out on the Renderosity Maya forum, when time permits.
February 18, 2013
Please note: If you find the color of the text hard to read, please click on "Printer-friendly" and black text will appear on a white background.