A few months back, we benchmarked the current crop of workstation graphics cards in some of the latest titles, just for kicks (How Well Do Workstation Graphics Cards Play Games?). At the same time, we were in the process of putting the latest FirePro and Quadro products through our professional graphics suite, along with a number of desktop-oriented gaming boards. Well, after literally several hundred hours of benchmarking, we have the data to go along with that follow-up story (and the results that go into our Workstation Graphics 2013 Charts).
Competing Graphics Cards Overview
Our field of contenders includes all of the heavy hitters. We have Nvidia's flagship Quadro 6000, as well as AMD's FirePro W9000, though our focus is more on the sub-$1000 category, since that's more in-line with practical budgets, even in the professional space.
A lot of readers requested that we also include desktop-oriented cards to see how they compare in workstation-class applications, so we added seven of those, too. It's actually interesting to track their performance in workloads like rendering, 2D drawing, and CAD with DirectX graphics output.
Here's a list of all of the cards we benchmarked:
| Nvidia | AMD | |
|---|---|---|
| Workstation (Current Generation) | Quadro K5000 Quadro K4000 | FirePro W9000 FirePro W8000 FirePro W7000 FirePro W5000 |
| Workstation (Previous Generation) | Quadro 6000 Quadro 5000 Quadro 4000 Quadro 2000 | FirePro V7900 FirePro V5900 FirePro V4900 FirePro V3900 |
| Gaming (Current Generation) | GeForce GTX 690 GeForce GTX Titan GeForce GTX 680 | Radeon HD 7990 Radeon HD 7970 GHz Edition |
| Gaming (Previous Generation) | GeForce GTX 580 | Radeon HD 6970 |
What We Couldn’t And Wouldn’t Include
AMD wasn't able to send over a FirePro W600 for our comparison. Interestingly, the company was willing to send over a FirePro S10000. This is a shame, since we wanted to dedicate analysis to the W600, since Nvidia doesn’t offer anything even remotely like it. A single-slot graphics card that can drive six monitors or projectors at the same time, and can even output six different audio streams would have been worth the effort, we think. Meanwhile, the FirePro S10000 card mentioned above, as well as Nvidia’s Tesla cards, are just too big for this story, though we do have a piece in the works covering Tesla. We also didn't include Nvidia's smaller Quadro 400 or 600, since they would have taken forever in some of our benchmarks, and wouldn’t have generated useable results in others due to their very limited performance.
Our Benchmark System
Choosing the supporting hardware to benchmark with put us in a bit of a quandary. The dual-CPU workstation based on a pair of AMD Opteron 4284 processors (Valencia, 3.0 GHz with Turbo Core) that we used last year turned out to bottleneck the higher-end graphics cards in this round-up. The same went for a system with two Intel Xeon E5-1660 processors running at 3.3 GHz. Under both platforms, the top-end boards we benchmarked landed very close together in the SPECapc tests, along with apps that didn't benefit from heavy threading, but instead needed big instruction-per-clock throughput from the host processor. Differences between the lower-end cards were much smaller, indicating the bottleneck.
We finally settled on a desktop-oriented processor with a high clock rate and fast memory. This might not reflect the average workstation very well (though Intel does sell a Xeon E3 configured very similarly that's often used in workstations), but it does give us a less constrained look at the differences between high-end graphics cards, some of which went from almost identical performance to huge performance deltas after transitioning to the new machine. In the end, the fastest cards sped up almost 20 percent in my new benchmark system than in the old workstation.
| Benchmark System | |
|---|---|
| CPU | Intel Core i7-3770K (Ivy Bridge), Four Cores, Eight Threads, 8 MB Shared L3 Cache, Overclocked to 4.5 GHz (Water Cooled) |
| RAM | 32 GB Corsair Dominator Platinum at 2133 MT/s |
| Motherboard | Gigabyte G1 Sniper 3 Rev 1.0, Z77 Express, BIOS F7 |
| SSD | 2 x Corsair Neutron 480 GB |
| OS | Windows 7 Ultimate x64 |
| Driver | Depending on Application |
Our Driver Selection
Driver selection is easy when it comes to benchmarking desktop graphics cards in games. Generally, we simply use the latest driver. Increasingly, this involves using beta software, since AMD and Nvidia are both trying to stay ahead of each other and the latest game titles before their software can be WHQL-certified.
This is done differently in the workstation space, and our round-up reflects that fact. Professionals pay more for these graphics cards because they need drivers certified by application vendors, guaranteeing the best possible reliability. You won't always get the best possible performance, but the certification does ensure proper functionality. Naturally, that's imperative, since any error can cost a company far more than the price of the hardware. Stability trumps performance in this space.
The same narrative applies to AMD and Nvidia equally, so our results don’t just speak to the theoretical performance of each GPU, but also to both companies' driver optimization. That can be more important than the raw potential of a card, creating interesting benchmark results that wouldn't seem to make sense otherwise.

Our Benchmark Software
We believe in a healthy blend of synthetic tests (for isolating specific performance attributes) and real-world application benchmarks. It’s also important to make sure that the selection of metrics is balanced so that the results don't get skewed toward one card vendor or the other. Finally, our selection of professional applications is limited to those for which we're able to attain legal licensing. I tried to include all of the titles that our readers wanted to see, and was mostly successful. There are some exceptions, though. For example, V-Ray’s makers never got back to me, so we aren't including their software. Also, some applications check the graphics hardware and simply won’t work with desktop cards.
Otherwise, we're only limited by time. Testing one card takes up to about 10 hours, which is to say an entire day. Clearly, putting this piece together was a labor of love.
A quick word on reading the graphs: the red bars always represent AMD's FirePro cards, and the green ones are always Nvidia Quadros. The black bars are reserved for the desktop graphics cards. We didn’t use different colors for the Radeons and GeForces because they mostly amount to a footnote when they're compared in professional applications.
Drivers Used
Interestingly, Autodesk has only certified drivers for FirePro cards that were originally released in 2010 and are now two generations old. Nvidia doesn't fare much better, with only the Fermi-based boards up to the Quadro 5000, as well as several older Tesla models, receiving any validation. Of course, even the newer cards run flawlessly with current drivers, since it was the hardware, and thus the cards themselves that were certified. Nonetheless, this patchy and intermittent state of certification on both sides of the aisle is both confusing and irritating for consumers.
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 310.90 (Certified) | Catalyst Pro 9.003.3.3 (Not Certified) |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 |
AutoCAD 2013: Overall 2D Performance
We're kicking off testing with the versatile and popular AutoCAD 2013 by Autodesk, first looking at its “2D” performance. The 2D is in quotation marks because AutoCAD 2013 actually uses D3D (DirectX) for its 2D and 3D output. This is doubly interesting because Windows hasn’t had universally-usable hardware acceleration through the kernel-mode driver since Vista, and graphics cards based on unified shader architectures don't have 2D units.
Unfortunately, the hardware implementation and application-specific optimizations for OpenGL are very operating system-dependent, and are stamped out with each change of the Windows driver model. This was one of the main reasons I decided against Windows 8 for this round-up, giving AutoCAD with DirectX a chance. Let’s take a look at the overall results before we start breaking down the individual outcomes.

The only observation that stands out is the lack of any card standing out. Every board demonstrates almost identical performance, meaning that any graphics card you buy, including gaming products, is usable for drawing in 2D. Most of the workstation cards do offer 10-bit color depth, but this is inconsequential for 2D tasks. The need for ECC-capable memory and the protection it offers from data corruption is also questionably in 2D workloads, so long as there aren't any additional compute tasks going on.
As expected, the four Cadalyst 2012 benchmarks paint a very similar picture to the overall chart on the previous page. The professional and desktop cards from Nvidia tend to inch out the AMD competition. But the margin is so small that every card boasts close to the same performance. The take-home message isn’t that Nvidia is faster, but that anyone who only uses AutoCAD for 2D tasks can use a desktop card and doesn't need to pay extra for a pro board.




Drivers Used
The driver situation isn’t any different for 3D in AutoCAD than it was for 2D. It’s regrettable that Nvidia stopped publishing specially-optimized drivers since the introduction of AutoCAD 2013 and its switch to DirectX.
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 310.90 (Certified) | Catalyst Pro 9.003.3.3 (Not Certified) |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
AutoCAD 2013: Overall 3D Performance
It’s easy to see that Nvidia's graphics cards have the upper hand once again. Even SLI provides a bit of a benefit, allowing GeForce GTX 690 to inch out the Titan. AMD drops the ball with its Radeon HD 7990, which gets beaten by the Radeon HD 7970 and performs at the same low level as an older Nvidia Quadro 2000. With few exceptions, the results are reminiscent of the usual game benchmarks.

The overall results are again based on four separate Cadalyst benchmarks. In particular, the Rotate Wireframe and Rotate Realistic benchmarks contribute a lot to the average, representing the two most important metrics of practical use. Any reasonably fast gaming graphics card should be ample for anyone limiting their professional use to AutoCAD.




Drivers Used
The driver situation in Inventor is identical to what we saw in AutoCAD 2013, since both applications are part of the AutoCAD Design Suite 2013 Premium.
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 310.90 (Certified) | Catalyst Pro 9.003.3.3 (Not Certified) |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Autodesk Inventor 2013: 3D Performance
Autodesk Inventor made it into our benchmark suite due to popular demand. We're using the well-known 1000-dice project, rotating, zooming in and out, and moving the view through a keyboard macro. Fraps is used to measure the frame rates.
The gaming graphics cards beat each company's workstation boards by a significant margin. Nvidia's GeForce GTX Titan doesn’t do well at all, ending up behind the Radeon HD 7970 GHz Edition. AMD's FirePro cards also excel in this benchmark, which is probably due to the many hidden surfaces involved in the workload (generally considered a strength of AMD's architecture).

Drivers Used
Nvidia’s 311.50 driver is a really good all-purpose build, but the most recently-certified driver when we ran all of our benchmarks was the 305.93, which is why we used it. Early in July, Nvidia published its 320.49 package, but there is no mention of changes to Maya's performance in the release notes. Regardless, the situation is a lot better than AMD's side of the fence, where searching for a Maya 2013-certified driver yields zero results.
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 305.93 (Certified) | Catalyst Pro 9.003.3.3 (Not Certified) |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Maya 2013: OpenGL Performance
For this specific page, we skipped Viewport 2.0 with DirectX 11 and focused on the older SPECapc 2009 suite because it employs OpenGL. I'm only interested in four tests at this point, and we'll come back to the photo-realistic rendering later.
The completion time for each benchmark is directly tied to each card's performance. And as we can see, the desktop cards don't stand a chance, regardless of how well they do under DirectX. Of course, this is a result of the drivers that accompany those workstation-class cards, which include optimizations that let them hardware-accelerate a great many functions in spite of the Windows Display Driver Model (WDDM). The gaming cards have to make do with a universal software emulation instead.
Nvidia's older GeForce GTX 580 posts some strange results. It manages to win one test by a wide margin and lose another one by a similarly large delta. Otherwise, the Nvidia and AMD cards trade blows, while SLI and CrossFire go to waste.




Drivers Used
When we get to SolidWorks, our driver situation improves. AMD recently put a lot of effort into fixing up its software, and it shows. Nvidia certified its 311.50 driver package before we started our testing (though again, it now has a newer 320.49 ODE build available), and that bolstered its SolidWorks performance by a massive 120 percent compared to the 310.90 version, vaporizing AMD’s marketing materials in the process. Consequently, what we're showing is quite a bit different from what AMD advertises in this test.
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 (Certified) | Catalyst Pro 9.003.3.3 (Certified) |
| Gaming | Not Supported | Not Supported |
SolidWorks 2013: Composite Score
SolidWorks 2013 is limited to workstation-class graphics cards. Unfortunately, the drivers we're using won't install on gaming boards, so we cannot include them. Additionally, if the software is run with non-certified drivers, there's supposed to be a quantifiable performance hit. The only exception is the version used by SPECapc 2013, which supposedly allows full use of SolidWorks 2013 whether the driver you're running is certified or not. We didn't bother testing, but rather used the certified drivers for this story.
First, we'll have a look at the graphics composite that both AMD and Nvidia like to use in their marketing materials.

AMD’s FirePro W7000 manages to inch out its direct competition from Nvidia, the Quadro 4000 and K4000, in spite of Nvidia’s newer drivers. The Quadro 2000 loses big against the entry-level AMD FirePro W5000. Strangely, though, the FirePro W9000 is beaten by Nvidia's older Quadro 6000, and not by a small margin. Nvidia's high-end card holds onto its crown in this benchmark.
Keep price and usage in mind when you look at the benchmark results, though. AMD's FirePro W9000 and W8000 aren’t simple polygon-pushers. They're designed for general-purpose computing. At the same time, Nvidia's Quadro 6000 is going to be more appropriate in the design department at a large company like BMW than in an engineer’s home office.
Once again, we have the four separate benchmarks, which give us an interesting look at driver performance. The comparisons between Nvidia's Quadro 6000 and AMD's FirePro W9000, as well as the Quadro K4000 and FirePro W7000, demonstrate that Nvidia manages to get the same or more performance from less powerful hardware.
We'll be taking a look at this in the pages to come once we get to our efficiency measurements. That'll give us more perspective on how AMD's FirePro W7000 and Nvidia's Quadro K4000 stack up in a power consumption test while running SolidWorks 2013. Overall, AMD is doing a pretty good job judging by the results so far and the graphics cards’ prices.




A Quick Introduction to SPECviewperf 11
Next up, we have the individual benchmarks from the SPECviewperf 11 suite. What makes this collection of tests special is that it employs the original source code from the applications it comprises. Consequently, we end up with a proper representation from a number of popular titles. There's one big drawback, though. SPECviewperf 11 uses old versions of those applications. Usually, that's still good enough to give us a general overview of an app's performance, but there are exceptions. Software that recently underwent major changes to its foundation won't be represented well at all. As an example, we already benchmarked Maya 2013 separately, allowing us to compare the results from an older engine on the next page.
Drivers Used
We used Nvidia's 311.35 driver in viewperf because the 311.50 build came out after we were already partly done with testing. At this point, though, with 320.49 out as of early July, it's possible (though unlikely) that the performance profile of these older titles is different. We really do love to have the very latest drivers in all of our stories (which is why you see Angelini constantly starting from scratch in his own benchmarks, such as in GeForce GTX 760 Review: GK104 Shows Up [And Off] At $250). This time, however, we have to accept that both companies continue working on their software, even as we're forced to draw a line in the sand and get busy with benchmarks.
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
CATIA (catia-03)
Nvidia dominates this benchmark, and the consumer graphics cards don’t stand a chance due to their missing driver optimizations. It’s very apparent that AMD’s Catalyst Pro driver isn't tuned for this benchmark or, more specifically, the engine on which it's based. The Quadro K4000 solidly beats the FirePro W7000 by a wide margin.

EnSight (ensight-04)
The consumer graphics cards fare more favorably in an older version of EnSight. The newer Quadro K5000 leads the pack by a mile. The more interesting (read: affordable) professional cards, AMD's FirePro W7000 outmaneuvers Nvidia's Quadro K4000.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
LightWave (light01)
Nvidia did its homework. The Quadro K5000 and K4000 are in front with a comfortable lead. Most of the other workstation graphics cards demonstrate very similar performance. The desktop cards suffer a resounding loss. The massive lead that Nvidia’s Quadro K5000 enjoys over the company's GeForce GTX 680, which should dominate the K5000 if you go by technical specifications, shows just how much the consumer cards are held back by their drivers.

Maya (maya-03)
The SPECviewperf 11 Maya benchmark yields similar results to the standalone Maya 2013 test we ran earlier. Once again, Nvidia’s Kepler-based Quadro cards take the lead, followed by their respective non-K counterparts. AMD's graphics cards follow with decent, but certainly not winning, performance. All of the company's newer professional cards perform about the same.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Pro/ENGINEER (proe-05)
This benchmark demonstrates, once again, that performance is almost completely dependent on optimized workstation-class drivers. All of the desktop boards get crushed. AMD's FirePro and Nvidia's Quadro professional cards are clearly better choices.
At the bottom of the chart, the GeForce GTX Titan gets stomped by AMD’s Radeon HD 7970 GHz Edition and finds itself in the back. Otherwise, all of Nvidia’s Quadro workstation boards, from the bottom to the top, finish ahead of the entire FirePro line-up. AMD’s cards finish very close to each other, suggesting a lack of proper driver optimization.

SolidWorks (sw-03)
Looking at the colors in the bar graph below, it's clear why the new version of SolidWorks doesn’t even let you use desktop graphics cards any more. There is no point, and the experience is miserable. AMD's FirePro cards perform interestingly, too. The new alliance between AMD and Dassault is really having an impact on driver development.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Siemens Teamcenter Visualization Mockup (tvcis-02)
Nvidia’s professional offerings are on top again, followed by AMD’s FirePro cards in the middle and the desktop pack in the back. As in Pro/ENGINEER, business-class software fares much better when you match it up to professional hardware and drivers. The gaming cards aren't even close to useable, except for maybe the Radeon HD 7970, which is able to beat one lone FirePro board.

Siemens NX (snx-01)
A similar workload yields similar benchmark results. This time, the Radeon HD 7970 GHz Edition is an astounding three times faster than Nvidia's GeForce GTX Titan. AMD's FirePro cards do a bit better here than in the previous benchmark.

The Unigine benchmarks are popular for their ability to help assess gaming performance in a synthetic DirectX 11 test. We use the OpenGL render path instead of DirectX, though, to see how these workstation-class cards do.
Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Unigine Heaven 4.0
Unigine Heaven's claim to fame started with the introduction of DirectX 11, and it quickly became a standard benchmark for evaluating tessellation performance. Currently it's in version 4.0, emphasizing the following features (taken from Unigine’s website):
- Comprehensive use of hardware tessellation, with adjustable settings
- Dynamic sky with volumetric clouds and tweakable day/night cycle
- Real-time global illumination and screen-space ambient occlusion
- Cinematic and interactive fly/walk-through camera modes
That the gaming cards do well in Heaven doesn't come as much of a surprise, despite the OpenGL render path. The workstation cards can’t lean on their driver optimizations to push them ahead, effectively leveling the playing field. Looking at cards like the FirePro W7000 and Quadro K4000, it becomes very clear how far ahead AMD's hardware is compared to Nvidia's, and just how much Nvidia is able to do with its drivers.

Unigine Sanctuary
The second Unigine benchmark puts its emphasis on somewhat different features (taken from Unigine’s website):
- Five dynamic lights
- HDR rendering
- Parallax occlusion mapping
- Ambient occlusion mapping
- Translucence
- Volumetric light and fog
- Particle systems
- Post-processing
These benchmark results are very similar to what we say in Heaven 4.0. One notable difference is that the top two graphics cards trade places.

Unigine Tropics
The last Unigine benchmark still sports OpenGL and lots of bright colors, but emphasizes different features. Again, taken from Unigine’s website:
- Dynamic sky with light scattering
- Live water with a surf zone and caustics
- Special materials for vegetation
- HDR rendering
- Parallel split shadow map
- Depth of field
- Real-time ambient occlusion
- Up to 2M polygons per frame
- Simulation of changing light conditions
In spite of the somewhat older engine and different emphasis, we’re faced with a familiar picture. Even though the FirePro W7000’s direct competitor in the professional space is Nvidia's Quadro K4000, the three Unigine benchmarks place it closer to the Quadro K5000. Hopefully AMD is paying attention to how much performance potential it's missing out on as a result of its drivers. The W7000 clearly plays host to more horsepower than some of the professional applications (or its price) would have us believe. This is a missed opportunity for AMD.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.5 Beta 2 |
GPU Caps Viewer PostFX
This benchmark was originally part of Nvidia’s software development kit (SDK). It combines 3D display via OpenGL with post-processing via OpenCL. The compute workload explains why AMD's Radeon HD 7970 GHz Edition, Radeon HD 7990, and GCN-based FirePro cards enjoy such a commanding lead. The benchmark doesn’t support CrossFire or SLI, and a manually-created profile doesn't improve the performance of either dual-GPU card.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.5 Beta 2 |
TessMark
As its name implies, TessMark benchmarks tessellation performance using OpenGL 4.0. The GeForce GTX Titan wins this metric by a wide margin in front of the second-place GTX 680.
Interestingly, AMD's workstation cards fare a lot worse than the company's gaming products. Tessellation apparently doesn’t have a place in AMD’s FirePro universe, and that's probably related to drivers.
It wouldn't be a stretch to suggest that this test appears to favor Nvidia, though. That's always something to consider when evaluating synthetics, which aren't indicative of performance in any real-world workload, but might unfairly suggest one product's superiority over another's. We're pointing out the possible imbalance upfront.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Not Supported |
| Gaming | 314.22 | Not Supported |
iray Renderer + 3ds Max
Despite somewhat worse performance, we're using iray for this benchmark category, since we don't have a V-Ray license. The performance results shouldn't be impacted, though, because both applications essentially do the same thing. Nvidia's GeForce GTX Titan inches out the older GTX 580. This deserves respect. But it also reminds us how Nvidia's GeForce GTX 600-series was hobbled from the get-go in compute-oriented workloads. Due to their underlying architectures, both Kepler-based Quadro suffer the same fate. Interestingly, our overclocked Core i7-3770K fares well against Nvidia's Quadro 2000.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 310.90 | Not Supported |
| Gaming | 314.22 | Not Supported |
Blender Performance
The results in Blender are similar to what we saw in iray. The margins between the top graphics cards’ performance numbers are larger, which is a sign of better optimization. Once again, only the Nvidia Quadro 2000 loses to the overclocked Core i7 with Hyper-Threading enabled, juggling eight threads at a time.
As opposed to iray, Blender can use both of the Nvidia GeForce GTX 690’s GPUs for rendering. It consequently pulls ahead of Nvidia's GeForce GTX 580. Nevertheless, the GeForce GTX Titan is way ahead of everything else. The workstation graphics cards again show that they aren’t cut out for these sorts of tasks.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.35 | Not Supported |
| Gaming | 314.22 | Not Supported |
Octane Performance
We see similar benchmark results for a third time. Nvidia's GeForce GTX Titan dominates the field, and the older GeForce GTX 580 mocks the rest of the field vanishing in its rear-view mirror. Clearly there is no shame in running a GF110-based board if CUDA-based apps are important to you.
Our version of Octane can only use one of the GeForce GTX 690’s GPUs. It might have been a contender for the top spot otherwise.
We didn't have an option to use the Core i7, but based on prior results, we'd expect it to roughly match Nvidia's Quadro 2000.



Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 310.90 | Not Supported |
| Gaming | 314.22 | Not Supported |
FluidMark Performance
Nvidia's GeForce GTX 690 uses the processing power of two GK104 GPUs to take a first-place win. The GeForce GTX 680 and Titan follow with almost identical performance. Kepler-based Nvidia Quadro graphics cards dominate their predecessors, and the Quadro 2000 finishes dead last.

Some Notes about OpenCL
After running all of those CUDA-based benchmarks, we decided that it'd only be fair to include some employing OpenCL, too. Unfortunately, CUDA is a proprietary API not supported if your card doesn't have an Nvidia sticker. OpenCL is the first platform-independent alternative, and in many ways, it's actually similar. For a variety of reasons, some apps still only support Nvidia's API (likely to avoid overly-complex code, which can be a pain to maintain). Still, a number of new OpenCL-based implementations show that it’s possible to generate code supporting devices from many different vendors.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Bitmining with OpenCL
Before we dive into bitmining, be aware that this isn't a prodigious use of your electricity. For more on why, check out All About Bitcoin Mining: Road To Riches Or Fool's Gold?
With that said, we’re using GUIMiner as the graphical interface for CLMiner to test OpenCL performance. Neither the GeForce GTX Titan nor the GTX 680 come anywhere close to AMD's Radeon HD 7990 and 7970 GHz Edition, both of which also best the FirePro W7000.
In order to test the dual-GPU cards, we started separate instances of GUIMiner and added the results together. We won't do this for the remaining OpenCL-based benchmarks because, if support isn't built in to the software, it shouldn't be part of the outcome.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
LuxMark 2.0: "Sala" Scene
The same pattern emerges, though this time, the GeForce GTX Titan pulls even further ahead of the GeForce GTX 680. We see further evidence that Kepler's underwhelming performance isn't a result of stacked benchmarks or a bad API, but rather the fact that Nvidia didn't prioritize compete, whereas AMD's GCN is much more adept in this discipline. In fact, Titan's OpenCL performance is still quite a bit under where it theoretically should be. Could it be that Nvidia simply isn't optimizing for OpenCL, instead putting its resources into CUDA?

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
ratGPU: Built-In Benchmark
The ratGPU results are interesting in several different ways. To begin, Nvidia's older GeForce GTX 580 wins. In fact, the top two spots are taken by previous-gen architectures. ratGPU utilizes hardware differently than other applications (like LuxRender), which is bad news for AMD's GCN design. Of course, we could have generated more difference between the faster cards by using a scene with higher complexity, though this would have simultaneously punished the slower cards severely.


Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Folding@Home Benchmark
We used the current Folding@Home benchmark with OpenCL, making it easier to compare the results. At the time of testing, there were problems with the CUDA-based version and cards like Nvidia's GeForce GTX Titan, along with some of the newer workstation boards. In contrast, OpenCL just works.
Nvidia's GeForce GTX Titan pulls ahead, demonstrating that not all of the company's GPUs struggle with OpenCL, but mainly the cards built on GK104. That conclusion is further supported when the GTX 580 performs well.
Two instances of the benchmark would have been needed for the Radeon HD 7990 and GeForce GTX 690 to work properly. That would have been a lot messier and more difficult than it was in Bitminer, though.
Single-Precision


Double-Precision


Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Single-Precision Financial Analysis
Financial analysis is a very good fit for graphics cards and parallel processing. The workstation cards do lag behind the consumer cards, but entrusting important calculations like these to a card without ECC RAM is risky. Consequently, this is really the time for solutions like AMD's FirePro W9000, built with general-purpose computing in mind rather than pushing out as many polygons as possible. Indeed, financial analysis is one of the professional workloads where these cards shine.
When we use single-precision math, five AMD cards lead the field, followed by the GeForce GTX Titan and 690. The FirePro W7000 even beats Nvidia's GeForce GTX 680. AMD's FirePro W9000 and W8000 might not be the strongest contenders in the graphics tests, but they post great performance numbers in this metric.
Binomial Option Pricing (FP32)

Monte Carlo Option Pricing (FP32)

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Double-Precision Financial Analysis
More complex math proves to be too much for some of these cards. Nvidia's GeForce GTX 690 edges out the Titan in the first benchmark thanks to its two GPUs, but falls way behind in the Monte Carlo option pricing benchmark. The performance hit attributable to FP64 is massive, making it very clear where corners were cut to make GK104 a well-balanced gaming GPU. AMD's FirePro W7000 doesn’t really do any better. Its double-precision performance is a far cry from the two Tahiti-based cards, since it centers on Pitcairn.
Binomial Option Pricing (FP64)

Monte Carlo Option Pricing (FP64)

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Encryption and Decryption Performance
Another GPGPU area is encryption and decryption. AMD completely dominates the field; not even the GeForce GTX Titan can come close.


Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Basemark CL
When we add another synthetic benchmark to our suite, we want it to give us a broader picture of the tasks that developers might use OpenCL to accelerate. You may already be familiar with the two simulations and two mathematical functions that make up the next set of tests.
Fluid Simulation
AMD's Radeon HD 7970 GHz Edition and the older GeForce GTX 580 do well in this benchmark. Nvidia's GeForce GTX Titan and 680 fall far behind. Even the Quadro 6000 catches the far more powerful Titan board.
The GeForce GTX 580’s excellent performance shows that not all of Nvidia’s gaming cards struggle with this benchmark.

Wave Simulation

The tables turn a bit. Nvidia's GeForce GTX Titan almost manages to keep up with the more powerful FirePro and Radeon cards. At the end of the day, though, the FirePro W9000 enjoys a decisive victory over AMD's Radeon HD 7970 GHz Edition thanks to a superior memory subsystem (and in spite of the gaming card's higher core clock rate).

Julia Rendering
AMD's cards really stick it to Nvidia's line-up. The Radeon HD 7970 GHz Edition takes the crown, followed by a number of FirePro boards. The nearest GeForce card shows up in sixth place, while the highest-ranked Quadro is in the bottom half of this chart.

Mandelbulb Rendering
And the tables turn again. After AMD’s victory in the last benchmark, Nvidia has its turn to dominate. Modern GeForce cards are almost twice as fast as AMD's Radeon family when it comes to Mandelbulb rendering, and the Quadro cards dominate their FirePro competition.

Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Basemark CL
The simple image processing filters included in Basemark CL are pretty similar to the normal array of filters found in popular photo editing applications. Due to the similarity between the algorithms used for the filters, the synthetic benchmark results represent the graphics cards' performance in this area fairly well. AMD's Radeon and FirePro cards win with ease. Only the GeForce GTX Titan manages to poke its nose up into the field of AMD hardware.
Image Processing
Drivers Used
| Driver | Nvidia | AMD |
|---|---|---|
| Workstation | 311.50 | Catalyst Pro 9.003.3.3 |
| Gaming | 314.22 | Catalyst 13.3 Beta 3 Catalyst 13.5 Beta 2 (Radeon HD 7990) |
Basemark CL
Similar to its image processing filters, Basemark CL’s video processing algorithms are a lot like the ones you might expect from real-world applications servicing this segment. The GeForce cards do a lot better with video processing than they did with image processing. AMD's FirePro W9000 also fares well, taking the top spot in each test except for the sharpening benchmark, where Nvidia's GeForce GTX Titan pulls ahead.
Video Processing




Efficiency: Nvidia Quadro K4000 Versus AMD FirePro W7000
The Quadro K4000 and AMD FirePro W7000 are both aimed at the average workstation customer and, consequently, are supposed to be top sellers. Their prices are also very similar, so it's only natural for us to compare them directly. Interestingly, Nvidia states that the Quadro K4000 has a thermal design power (TDP) of 80 W, while AMD's FirePro W7000’s is rated at 150 W. The two cards perform about the same in an application like SolidWorks. Take that fact, along with the power numbers, and you might assume AMD's card uses twice as much power to do the same work. But the FirePro is actually a higher-performance card on paper. Let's take a look at a couple of different scenarios.
Scenario 1: Similar Performance in Real-World Workload with Partial GPU Load
Going back to SolidWorks, we can compare the power consumption of Nvidia's Quadro K4000 and AMD's FirePro W7000. The Pitcairn-based card uses more power under lower and peak loads.

The next question is how each card's power consumption relates to its graphics performance.



AMD's FirePro W7000 draws about 10 percent more power than the Quadro K4000 and provides five percent more performance in tasks where the two cards fare close to the same. This really isn’t so bad, since a 5 W difference isn’t anything to write home about.
Scenario 2: Different Performance in Real-World Workload with High GPU Load
I didn’t want to use a completely artificial GPGPU stress test to apply 100 percent load. Instead, I opted for the Unigine Heaven 4.0 benchmark, which scales well and uses almost 900 MB of memory. This doesn’t torture the cards as intensively, but it is more realistic than a straight-up compute workload pushing each card as hard as possible.
Interestingly, the two resulting lines aren't quite identical, even though each GPU is faced with the same task. For a better chance at fairness, the two cards were warmed up to operating temperatures by a medium load from Maya 2013’s Toy Store benchmark scene.

We might have received an exceptional FirePro W7000 press sample, but it looks like the Pitcairn-based card scales very well. That means its TDP is rated very conservatively in AMD’s technical specifications. Not even a full-out stress test takes the card anywhere close to 150 W. It barely breaks 100 W in a benchmark that caused two Malta cards to thermally throttle in Radeon HD 7990 In CrossFire: The Red Wedding Of Graphics.
The bars below show us that, even though the FirePro W7000 is less efficient than the Quadro K4000 under medium loads, the opposite is true under a more taxing scenario. The ratio of power consumption to graphics performance is clearly in AMD’s favor, even though this benchmark generally tends to favor AMD a bit.



We thought it'd be interesting to repeat the experiment with a normal Radeon HD 7850 and 7870, since the technical specifications put AMD's FirePro W7000 somewhere in between those two cards. Our FirePro sample draws about as much power as the frugal Radeon HD 7850, but it performs significantly better. This outcome is plausible. Unfortunately, we couldn't try the same comparison using SolidWorks, since the test won't start with a desktop card installed.
Bottom Line
The two examples we set up are frankly somewhat arbitrary. Nvidia's Quadro K4000 is twice as fast as the FirePro in some CAD applications, and OpenCL-based tests show the complete opposite. But the scenarios we picked are still very interesting for two reasons. First, they show how far good driver optimization can take a card. Second, they demonstrate what happens when a GPU is able to make actual use of all the theoretical performance it has on paper.
Conclusion
Even though the results differ massively from one benchmark to the next for both AMD and Nvidia graphics cards, there’s somewhat of a common thread. We'll cover our first impressions for each company separately, and then reveal the two cards we liked best.
Nvidia: Steady Performance or Slightly Improved, Efficiency Much Better
You have to give it to Nvidia: the company is squeezing a lot of performance out of its GPU hardware. Comparing the Quadro K5000 to the Quadro 5000 and the Quadro K4000 to the Quadro 4000 are two good examples. The K5000 even comes close to the Quadro 6000's performance on occasion, while using a little more than half as much power in the process.
However, the Quadro K5000 and K4000 share some of the limitations of Nvidia’s gaming graphics cards based on GK104. Simply, that GPU wasn't designed for compute. Nvidia’s drivers are still top-notch, and its cards are going to be your first choice in an application the company is optimizing for.
AMD: Performance Vastly Improved, Efficiency Steady or Slightly Improved
AMD deserves respect for its GCN-based FirePro cards. They offer vastly better performance than their predecessors. AMD caught up in many areas where it previously trailed Nvidia in the professional space. And if a workload overwhelms one of these boards, it's a driver issue, not a limitation of the hardware.
The company is also in a better place with its drivers for many applications. Of course, there is still room for improvement. The GCN-based cards naturally do well in compute-heavy applications via OpenCL support. There's a real alternative to Nvidia’s Quadro cards, particularly when you take price into account. Again, this is as long as AMD's driver is optimized for the workload in question. In titles that haven't received much attention yet, performance is less compelling.
Recommended Workhorses: Nvidia's Quadro K4000 and AMD's FirePro W7000
Nvidia's Quadro K4000 and AMD's FirePro W7000 simply offer the most bang for the buck. The pricier cards in both companies' higher-performing tiers are typically too expensive for the average professional, and often aren't the right choice anyway since they emphasize compute-intensive workloads. Pay particular attention to the types of software you plan to run on your workstation, and pick a professional graphics card accordingly. If your vendor of choice hasn't put much effort into optimizing for it, then there's a good chance you're going to be disappointed.
If the two cards we're most excited about are too expensive, then AMD's FirePro W5000 is a good alternative, delivering decent performance from a cut-back version of the Pitcairn GPU. Nvidia also sent along one of its Quadro K2000 cards after this story was completed, and we updated our Workstation Graphics 2013 Chartsto reflect that board's performance, too.
Bottom Line
For the most part, gaming graphics cards don't work for professional applications, and increasingly, ISVs are requiring workstation-class hardware. The only real exceptions are DirectX-based titles like AutoCAD 2013 and Inventor 2013, where the additional optimizations to a pro card and its drivers aren't necessary. There are also certain compute-heavy applications for which desktop-oriented cards perform well also, so long as you can live without features like ECC memory. But if one messed up byte could throw your result off, sending Wall Street into a tailspin, a workstation graphics card designed for the job is a smart choice.
When we look at the market as a whole, AMD is much more competitive now than ever before, while Nvidia continues to optimize and polish its existing products. The race hasn’t been this exciting in a long time. It remains to be seen if AMD can get its drivers certified for more applications. After all, the tremendous architecture that works so well for the company in the gaming space has a ton of potential in the workstation segment, too. Makes us wonder if an excellent software bundle might do wonders for its workstation line-up?






































