Friday, November 19, 2010

AMD FirePro V9800 4 GB: Eyefinity Meets Professional Graphics

What's red, black, sports 4 GB of GDDR5, and has DisplayPort outputs all over? AMD's FirePro V9800, the company's ultra-high-end workstation card. We run it through a battery of SPECapc and real-world performance tests, then compare it to Quadro 5000.
I remember sitting aboard the USS Hornet and listening to AMD pitch its Eyefinity technology for the first time. Back then, company representatives were so anxious to talk about Eyefinity’s potential for gaming. All of the sudden, we had this GPU that was powerful enough to drive an array of monitors at 5760x1080, no sweat. And two of those chips could be used together in CrossFire to realistically make 5760x2160 viable.

The problem, it turned out, wasn’t with AMD’s vision. Eyefinity across three displays remains an impressive sight for anyone used to gaming on just one screen. But the six-monitor stuff fell flat—an unfortunate consequence of the fat bezels nearly all LCDs continue to employ. More than a year after AMD’s first demonstrations of Eyefinity, we still haven’t seen Samsung’s MD230X6 six-monitor setup—though it’s available on Newegg for $3600 bucks. Good game.
As AMD talked about gaming on multiple displays, however, all I could think about was the four-monitor setup I had at home. Would adding two screens help improve productivity? Did I even have room on my desk for such a massive amalgamation of desktop real estate? Word. Excel. Outlook. Firefox. Six or seven Skype windows. I certainly had enough software open all day long to use six screens. That counted for something.
Professionals Can Use This Stuff
As a gaming-oriented card, the Radeon HD 5870 Eyefinity 6 Edition honestly struggles with its intended purpose—again, through no fault of AMD’s. As a productivity-oriented card, the Radeon HD 5870 Eyefinity 6 Edition is overkill. You don’t need a flagship GPU to drive a sextet of displays, after all. The good news is that DisplayPort 1.2 will make it possible for the new Radeon HD 6850 and 6870 boards to support six displays using hubs (once those are available, that is). There won’t be any need for a special edition card to enable Eyefinity’s full complement of outputs.
You like that? Even these little 21.5-inch Dells take up more than five feet of horizontal desk space.You like that? Even these little 21.5-inch Dells take up more than five feet of horizontal desk space.
More good news—in the workstation space, you don’t really need to worry about a card that might not please everyone. AMD knows that its new FirePro V9800 is only going to appeal to a very narrow sliver of the professional market. After all, the thing costs $3500 bucks, it sports a full 4 GB of GDDR5 memory, and it’s another one of those products specifically designed to support six monitors via DisplayPort. For those who can use the FirePro V9800, though, this board is the first of its kind, and it might just enable usage models and configurations that simply were not possible previously.
Back in September, Tom’s Hardware DE reviewed the $1500 FirePro V8800 against Nvidia’s Quadro 5000. The Quadro card performed significantly better—but we were using one of the first drivers available for AMD’s card. Since then, updates have been made, purportedly improving performance. With that in mind, if you don’t specifically need the massive frame buffer, six-display array, or frame lock/genlock support, then the FirePro V8800 is going to be a much more economical buy than the $3500 FirePro V9800.
Zoom
So, how does the FirePro V9800 (with the newest drivers) compare to Nvidia’s Quadro 5000? We fired up as many professional-class tests as we could get our hands on to find out. Oh, and I also set up an array of six Dell monitors to show you what working in After Effects, Premiere Pro, and Photoshop CS5—all at the same time—might look like.

We went into depth on the FirePro V8800 and Quadro 5000 back in September with Workstation Shootout: Nvidia Quadro 5000 Vs. ATI FirePro V8800.
For the most part, AMD’s FirePro V9800 is very similar to the V8800, with a handful of notable exceptions. Both cards employ AMD’s Cypress GPU with all 1600 stream processors enabled. The V8800’s core clock is set to 825 MHz, while the V9800 set to 850 MHz (that 25 MHz difference should be negligible in most benchmarks).
AMD's "High-End" LineupFirePro V9800FirePro V8800FirePro V7800
GPUCypressCypressCypress
Stream Processors160016001440
Single-Precision TFLOPS2.722.642.02
Memory4 GB GDDR52 GB GDDR52 GB GDDR5
Memory Bandwidth147.2 GB/s147.2 GB/s128 GB/s
I/O6 x DisplayPort
3-pin Stereo
Frame Lock/Genlock
4 x DisplayPort
3-pin Stereo
Frame Lock/Genlock
1 x DL-DVI
2 x DisplayPort
Frame Lock/Genlock
Form FactorDual-SlotDual-SlotSingle-Slot
Power199 W208 W138 W
MSRP$3499$1499$799

They also both sport GDDR5-based frame buffers running at the same 1150 MHz (yielding the same 147.2 GB/s of memory bandwidth on the same 256-bit bus). Granted, the V8800 includes 2 GB of memory, while the V9800 comes equipped with 4 GB. But don’t let twice the memory fool you—it’s actually pretty difficult to take advantage of that much RAM unless you’re working with scientific/oil/gas data sets or 4K frames in stereoscopic 3D.
Zoom
The two cards occupy dual-slot form factors and support CrossFire. But while the FirePro V8800 only needs two six-pin power connectors, the V9800 requires one six-pin and one eight-pin connector. Moreover, the V8800 offers a quartet of full-sized DisplayPort outputs and a three-pin stereo connector. The V9800 fully utilizes Eyefinity through six mini-DisplayPort connectors (the same stereo plug is available, too).
Eyefinity
Because the FirePro V9800 uses DisplayPort exclusively, there are bound to be questions about hooking up to DVI-based monitors. The good news is that AMD includes five DisplayPort-to-single-link DVI adapters and one DisplayPort-to-dual-link DVI adapter with the card. The single-link dongles support resolutions of up to 1920x1200, while the dual-link dongle enables 2560x1600 on one display.
Zoom
Of course, configuration is a lot simpler if you go with native DisplayPort monitors, like the Dell P2210Hs on my test bench. Connecting six of them is a simply matter of running mini-DisplayPort-to-DisplayPort cables between the components.
Perhaps best of all (especially for professional users), most of the DisplayPort-oriented bugs we encountered back in March with AMD’s Radeon HD 5870 Eyefinity 6 Edition card have seemingly been worked out. No more monitors disappearing from our Display Group. No more corrupt cursors persisting on-screen. And no more driver-reported failures. The only issue we encountered was a tendency for our 30” Dell 3007WFP to not receive a signal during boot-up (attached to a dual-link active adapter), forcing a manual reboot.

Test Hardware
ProcessorsIntel Core i7-980X (Gulftown) 3.33 GHz, LGA 1366, 6.4 GT/s QPI, 12 MB Shared L3, Hyper-Threading enabled, Power-savings enabled
MotherboardGigabyte X58A-UD5 (LGA 1366) Intel X58/ICH10R, BIOS FB
MemoryKingston 6 GB (3 x 2 GB) DDR3-2000, KHX2000C8D3T1K3/6GX @ 8-8-8-24 and 1.65 V
Hard DriveIntel SSDSA2M160G2GC 160 GB SATA 3Gb/s
GraphicsAMD FirePro V9800

Nvidia Quadro 5000
Power SupplyCooler Master UCP-1000 W
System Software And Drivers
Operating SystemWindows 7 Ultimate 64-bit
DirectX DirectX 11
Graphics DriverAMD Catalyst 8.773

Nvidia Quadro Release 259.57

Although AMD positions its FirePro V9800 against Nvidia's Quadro 6000, it doesn't take a 6000 to offer up competitive performance, we've found. If you're more interested in a comparison to the 6000, know that it includes more shader processors (448 vs. 352), more memory (6 GB vs. 2.5 GB), more memory bandwidth, and consumes more power. Priced at $5000, it's like the FirePro V9800 in that it appeals to a very tiny sliver of the market who can put its massively large frame buffer to use.
Zoom
Benchmarks and Settings
Video Encoding
BlenderCustom Workload; cd blender-2.54-beta-windows64, bench_blender.bat, rem blender -b thg.blend -o testset -f 1
Applications
Adobe After EffectsCS5; Custom Workload, SD project with three picture-in-picture frames, source video at 720p; Adobe Paladin workload (with hardware Mercury Playback Engine support on Quadro)
Adobe PhotoshopCS5; Custom Workload, Radial Blur, Shape Blur, Median, Polar Coordinates filters
Adobe Premiere ProCS5; Custom Workload, 1280x720p, 59.94 FPS video, Panasonic DVCPro100, HVX-200 camcorder on P2 media, Render to Work Area.
Adobe Media EncoderCustom Workload, Encode Premiere Pro project to h.264 for Blu-ray
e-on Software Vue 8 PLE1920x1080 landscape render, Global Illumination enabled
NewTek LightWave 3D 9.6Custom Workload, high-poly Tom's Hardware logo; clone surface detail across logo (Modeler, script); OpenGL preview of 600-frame animation of Tom's Hardware logo (Layout); 1.68 million polygon render of four captured frames using motion blur, ray-traced shadows, global illumination.
Autodesk 3ds Max 2010Default Scanline Renderer and Mental Ray Renderer, both test scenes from 3dspeedmachine.com
Autodesk MatchMover 2011Custom workload, 720p camera footage tracked in 3D space
Synthetic Benchmarks and Settings
SPECapc LightWave 9.6LightWave 3D Discovery Edition, Render and MT benchmark scores
SPECviewperf 11Default GUI options, 2560x1600; Workloads: CATIA, EnSight, LightWave, Maya, Pro/E, SolidWorks, Teamcenter Visualization Mockup, NX
SPECapc Maya 2009Default run, composite score generated through SPEC Excel template
SPECapc 3ds Max 9Default run, composite score generated through SPEC Excel template

SPECviewperf 11AMD FirePro V9800Nvidia Quadro 5000
catia-0315.0044.57
ensight-0424.3138.13
lightwave-0149.8665.53
maya-0364.2267.32
proe-054.8510.83
sw-0238.8861.27
tcvis-0223.4338.86
snx-0126.4939.31

A heads-up battle in SPECviewperf 11 seems pretty bad for the FirePro card, even with AMD’s latest drivers. Things actually look a lot better for AMD than they did in our FirePro V8800 review, true. But Nvidia still walks away with a victory in each of the disciplines tested.
SPECapc LightWave 9.6

The interactive and render tests are run sequentially, generating both scores. The MT (multi-tasking) test sees the interactive and rendering tests executing concurrently.
Nvidia’s Quadro 5000 takes a small lead in the render and multi-tasking test, but the interactive benchmark favors AMD’s solution—something we saw in Uwe’s FirePro V8800 review as well.
SPECapc 3ds Max 9
SPECapc 3ds Max 9AMD FirePro V9800Nvidia Quadro 5000
Wireframe Graphics2.092.04
Mixed Wire/Shade GFX2.892.98
Shaded Graphics4.334.53
Hardware Shaders10.1110.36
Graphics, Texturing, Lighting, and Blending2.423.74
Inverse Kinematics2.963.12
Object Creation, Editing, and Manipulation4.004.03
Scene Creation Manipulation4.523.75
Rendering14.7614.92


The CPU render and hardware shader tests are close, even if each favors Nvidia. The graphics benchmark is much more pronounced, going in favor of Nvidia’s Quadro 5000 card.
SPECapc Maya 2009

Using the latest drivers, AMD’s FirePro V9800 was unable to finish the entire SPECapc Maya 2009 suite, returning zeros for the latter portion of the test, which resulted in the CPU and I/O portions failing. Thus, all we have here is a graphics score. The good news is that AMD’s score is much higher than Nvidia’s. The bad news is that there’s no overall composite score to compare, since the FirePro doesn’t complete the test.
Now, the challenge with any of the SPEC tests is that they’re based on old versions of applications that get updated every single year. None of the latest trends in software development get taken advantage of, potentially leaving performance on the table. Unfortunately, the member organizations that make up SPEC seem to move slowly (gasp—bureaucracies are horribly inefficient?), so the snapshots of workstation graphics come with a time delay.
We were marginally successful getting a handful of other workstation-class apps running, but were amazed to find how many of these tasks are CPU-limited…


Our standard Photoshop benchmark is optimized for threading, but it’s not built to test the few OpenGL-accelerated tasks Photoshop can execute. Thus, it’s hardly surprising to see both graphics cards turning back identical performance.

Our After Effects project doesn’t make extensive use of the Mercury Playback Engine. Nevertheless, Nvidia gets the job done just a little faster.

Our render job finishes faster on Nvidia’s Quadro 5000, but the Adobe Media Encoder task is, fundamentally, just as fast on both graphics cards. Again, this is an entry-level rendering job, though. What happens when we crank things up from weaksauce to hot sauce?

Paladin, directed by John Allardice, was distributed by Adobe during the CS5 launch to showcase what the software can do. It’s much more effects-heavy than our in-house benchmark, setting it up as a wonderful worst-case scenario for Hollywood-quality rendering jobs.
The impact of GPU-based acceleration here is phenomenal. Premiere Pro 5.0.2 added support for the Quadro 5000, but you don’t necessarily need one of the supported cards to accelerate the Mercury Playback Engine. Check out William Van Winkle’s story on Adobe CS5 for more on enabling hardware support with any CUDA-enabled GPU.

In addition to the SPECapc tests, we wanted to take 3ds Max and LightWave for a spin using real-world workloads.
Whereas the SPEC tests take a variety of measurements and wrap them into composite scores, we simply timed two tests. One employs the default scanline renderer and the other uses the mental ray renderer.
Unfortunately, neither benefits from GPU-based acceleration, like the Arion light simulator (optimized for CUDA). The result is two very similar scores across both cards.


The OpenGL-based LightWave preview finishes just a tad quicker on the FirePro card, while the very CPU-dependent Modeler workload finishes in the same amount of time.
Zoom
From there, we rendered four individual frames from our Tom's Hardware logo test:
LightWave 9.6 Frame RenderAMD FirePro V9800Nvidia Quadro 5000
Frame #89:459:44
Frame #4110:1110:11
Frame #50011:2011:19
Frame #6008:088:07

Again, we're looking at a processor-bound scenario. It all goes to show that, even in a workstation environment, building with balance in mind is a smart decision. Want more proof? Check out the Xeon 5600-based benchmarks that demonstrate how well LightWave rendering scales with as many as 24 threads.

I have to believe that power consumption is less of a consideration in the ultra-high-end workstation space than any other market. If you need one of these card’s muscle, then you need it. Even 100 extra watts isn’t going to dissuade you if a certain graphics processor will help get jobs done faster, allowing you to make more money.
With that said, the results here are hardly a surprise. AMD’s FirePro V9800 enjoys slightly lower idle power use—a testament to the company’s ability to spin down its Cypress GPU when it’s not being used. Under load, however, you can’t get away from the fact that this board hosts a massive 4 GB frame buffer and the architecture’s complete implementation. Nvidia benefits from a smaller frame buffer and a significantly cut-back version of the GF100 GPU.

When Uwe reviewed the FirePro V8800, he concluded that it trailed the Quadro 5000 in almost every benchmark and rightly awarded Nvidia’s board a victory.
Then AMD got its drivers into better shape. The FirePro V8800 would likely still trail the Quadro by a notable margin today. Remember that the FirePro card also sells for significantly less, though.
Today’s comparison is far less balanced from a pricing perspective. The V9800 is a $3500 card, while the Quadro is available under $2000. That premium gets you a 4 GB frame buffer (which AMD says is good for massive rendering tasks), six display outputs (which works well for professionals doing their jobs across huge desktops), and frame lock/genlock support. If none of those features sound important to you, the FirePro V8800 is a more sensible board to consider. Fortunately, the performance reported here, with the latest drivers, should be representative of that card, too. The V9800 only runs 25 MHz faster than the V8800, and we didn’t use any workloads capable of taxing that large repository of GDDR5. From that angle, we’re simply updating Uwe’s results from September with newer software, and coming to the same conclusion—generally, Nvidia’s Quadro is the faster card.
AMD and Nvidia each have their own fortes that make generalizations much less meaningful than they’d be in a desktop graphics card review.
If you’re a creative professional working with Adobe’s CS5 suite, then the Quadro is hands-down a no-brainer. The fact that Nvidia had the engineers to help develop Adobe’s Mercury Playback Engine was a huge strategic win for both Adobe and Nvidia. The former was languishing in its development, only just implementing 64-bit support, while the latter needed a more complete GPU-based acceleration solution than Elemental’s Accelerator plug-in for CS4. The extent to which hardware support speeds up the rendering of effects-heavy work simply embarrasses the FirePro lineup. And if you didn’t watch the video of the Paladin trailer we used for testing, you really owe it to yourself to check out.
At the same time, it’s possible that Nvidia’s emphasis on GPGPU computing caused it to fall behind elsewhere. AMD’s Eyefinity technology has proven to be a boon on the desktop space, and it’s now strutting its stuff in the workstation market. The FirePro V9800’s ability to drive six DisplayPort-equipped displays concurrently is unprecedented. If that’s a capability that means something to you, you’re only going to find it in one place. Even AMD’s FirePro V8800 can do four simultaneous display outputs, while Nvidia’s highest-end board is limited to a now-pedestrian two.
Clearly, this story doesn’t end with a handful of tests from SPEC. There was a period there, after the workstation graphics market consolidated down to ATI and Nvidia, that professional cards looked a lot like their desktop counterparts with some special driver sauce sprinkled on top. Now we’re looking at true differentiating features that steer professionals toward one product or the other based on their tasks.

No comments:

Post a Comment