NVIDIA Quadro K5200 - In review
November 2, 2014 10:23 am
NVIDIA upgraded their Quadro workstation graphics card lineup. Among the new members are the new Quadro K5200, which replaces the Quadro K5000. I was sent a card to test drive for a while and see what I thought. This review is the result.
In short, it's super-great. Really, it is. Image courtesy NVIDIA.
The Quadro brand implies "workstation" class cards. This means they are built for using workstation applications, i.e. Autodesk Maya, Autodesk 3ds Max, Autodesk Inventor, SolidWorks - you know digital content creation, engineering and CAD type software. They're also great cards for many scientific, economic and engineering problems because of the vast number of general purpose CUDA computing cores and the large memory capacities the cards offer. Did I mention medical and energy exploration uses too? Yeah. Throw those in. Basically, whenever you get an MRI at the hospital and Dr.Spaceman is examining some dense voxel data on screen of your busted kneecap - they're using an NVIDIA Quadro card. They're everywhere.
Graphics cards have become far more than just graphics processing units. They're also a general purpose, massively parallel compute device. They're supercomputers on a card, especially the Quadro lineup. You can write code to run in parallel on all 2304 compute cores (in the case of the Quadro 5200) using either CUDA or the more vendor neutral OpenCL. This is why things like Pixar OpenSubdiv libraries, now found in the latest version of Autodesk Maya, and certain tools in the latest version of the Adobe suite, or the compositor Nuke by The Foundry are so ridiculously fast; They're embarrassingly parallel problems that have been implemented in CUDA or OpenCL and run on all cores en mass. It's classic divide and conquer.
This review will use industry standard benchmarks to evaluate the Quadro K5200 and how it compares to its predecessor, the Quadro K5000. I'll also throw in my own subjective experience with them while using Autodesk Maya and Autodesk Mudbox. Finally, be sure to check out my other review of this card's little brother, the Quadro K2200 (coming soon).
In the words of NVIDIA, the new Quadro K5200 is:
"Extreme performance to handle demanding workloads, 8 GB of GPU memory, advanced four-display capabilities for large-scale visualization, and support for high-performance video I/O."
The Quadro K5200 replaces the older Quadro K5000. It makes some marked improvements over its predecessor. A basic comparison:
|Feature||Quadro K5000||Quadro K5200|
|Core clock||706 MHz||650 MHz|
|Memory clock||1350 MHz||1500 MHz|
|Memory bus bandwidth||173 GB/s||192 GB/s|
|Memory size||4 GB||8 GB|
|Peak power consumption||122 watts||150 watts|
More computing cores! The Quadro K5000 had 1536 CUDA cores, the new Quadro K5200 increases this to 2304 CUDA cores - that's an increase of 768 extra cores, or about 67%. The GPU itself has also been changed from the GK104GL to the more powerful GK180GL - the same GPU used in the Quadro flagship card, the Quadro K6000.
The OpenSubdiv library, used by Autodesk Maya 2015, leverages NVIDIA CUDA and OpenCL to handle massive data sets interactively. This is a screen shot of the interactive viewport - and it was indeed interactive, easily maintaining more than 60 frames per second. The heads-up-display only shows the polygon count as Maya generated polygons. The actual polycount taking the OpenSubdiv displacements into account is in the tens of millions.
Memory capacity has doubled from 4GB of GDDR5 to 8GB of GDDR5. The memory bandwidth also increased from 173 GB/s in the Quadro K5000 to 192 GB/s in the new Quadro K5200. Memory clock also received a small upgrade from 1350 MHz in the Quadro K5000 to 1500 MHz in the new Quadro K5200.
Unfortunately, the performance increase also caused an increase in peak power consumption. Peak power consumption increased from 122 watts with the Quadro K5000 to 150 watts with the new Quadro K5200. That's a power usage increase of 25%, but the average overall performance increased by roughly double that. In all, a good tradeoff.
The new Quadro K5200 features the same connectors as its predecessor. The physical form factor also remains the same. The Quadro K5200, like it's predecessor, is 10.5" long and takes up two slots. The cards are exactly the same size. If a Quadro K5000 fit in your case, the new Quadro K5200 will fit as well.
Note the longer PCB in the new Quadro K5200 as compared to its predecessor the Quadro K5000. The new card's PCB extends the full length of the fan shroud. The card's overall footprint however remains identical. (Don't mind the dust on the K5000; I put it through its paces. Rock solid performer, as I expect the K5200 will be.)
Both the old and the new card feature a DVI-I DL, DVI-I SL and two DisplayPort v1.2 connectors, allowing a total of 4 displays to be connected at once. One difference, however, is the much larger and additional exhaust vents found on the newer Quadro K5200.
The new Quadro K5200 (top) has much larger air exhaust vents.
Why the larger air vents? After pulling off the plastic fan shroud, we can see another physical difference between the cards. The newer Quadro K5200 is equipped with a significantly larger heatsink to try and help dissipate those extra 768 CUDA cores.
The new Quadro K5200 (bottom) has much larger internal heatsink.
Despite this, there is no noticeable fan noise like we had with its great grandfather, the Quadro 5000 (sans the 'k'). As with the rest of the Kepler-based lineup, the new Quadro 5200 runs pretty cool. Contrast this with the older Fermi-based cards where we could just about fry an egg on the PCB. The shroud of the new Quadro K5200 rarely exceeds 130 °F (≈54 °C) in a well ventilated case. It's not only more computationally powerful than it's predecessors, it's also much more efficient.
I benchmarked the two cards in SPECviewperf® 11 as well as the newer SPECviewperf® 12. For those not familiar with it, SPECviewperf® is the workstation analog of the well-known 3DMark gaming card benchmark. It's pretty much the industry standard benchmark for workstation graphics cards. One must keep in mind, however, SPECviewperf® is a synthetic benchmark. The reported performance numbers will vary from one machine to another, even with identical graphics cards. This is not unexpected.
Keep in mind the best measure of performance is with real-world data that the end-user will actually be working with. What SPECviewperf® does allow us to do, however, is to compare multiple cards tested on the same machine.
We cannot compare SPECviewperf® 11 scores to SPECviewperf® 12 scores. The test suites use entirely different data sets. The scores do not correlate in any way.
First, lets look at the newer SPECviewperf® 12 results to see how the new Quadro K5200 stacks up to its predecessor, the Quadro K5000:
Indeed, there have been some marked improvements.
I also ran benchmarks against the older SPECviewperf® 11, on the same hardware as my previous Quadro benchmarks. This allows us to make direct comparisons between cards to see how performance has evolved across several Quadro generations. It also gives us an idea of what to expect, should we be looking at a used Quadro card:
Remember, we cannot compare these to SPECviewperf® 12 results. These results can only be compared within this graph to one another.
The subjective testing in Autodesk Mudbox revealed the new Quadro K5200 to be invincible. I completely saturated the 16GB of available memory on the test machine before I could ever get the interactive frame rate to drop below an acceptable level. Slowdown? What is slowdown? We don't have no stinkin' slowdown!
Things were pretty much awesome.
The NVIDIA Quadro K5200 is a superb card. Retailing for around $2,000 it ought to be. (You can find it slightly cheaper. Shop around.) It's much closer to the flagship, the Quadro K6000 (which retails for around $5,000) than its predecessor ever was. The increased number of cores and the upgrade from the GK104GL GPU to the GK180GL GPU, in addition to the 768 extra cores, boasts a significant improvement in overall compute performance. In terms of compute performance, it pretty much blows its predecessor away.
With constant advances in performance, dealing with massive data sets is rapidly becoming easier. This is even more important today than it was even two years ago. More and more algorithms are making use of the general purpose parallel computing facilities on these cards - be it OpenCL or CUDA.
Computer graphics is a very competitive market and having a great performing card translates directly into economics. If I can hardware render something in a fraction of the time it used to take, or work a data set twice as big, or sculpt at a much higher subdivision level than was previously possible - or deliver a shot to a client faster - that's a big deal. There will come a time when nobody remembers what "software rendering" was and you can bet the farm NVIDIA Quadro will be written all over it!
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.
November 3, 2014
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.
Please take a moment to join Renderosity's Newsletter List to receive news and information from Renderosity!