• CGSociety Hardware Review
    28 March 2007, by Joseph Tan
     
    {digg}
    NVIDIA has recently released a series of high-end Quadro FX workstation graphics cards. CGSociety was loaned a series of cards from a local supplier, LeadTek, and we will focus on the recently announced Quadro FX 4600. The new Quadro FX products Quadro FX 4600 and the 5600 are based upon the G80GL GPU and are aimed at high-end applications.

    The NVIDIA Quadro products are sold at a price premium compared to their consumer-oriented GeForce products. According to NVIDIA, the Quadro FX range are ISV certified with the leading professional applications, support commitment of three years from release, guaranteed availability for 18 months from release, Quadro memory management optimization, fast pixel read-back performance, and additional workstation-specific OpenGL acceleration support.
    Admittedly, every GeForce product from my memory has enjoyed updated driver support for at least three years after release, and are available for years from retailers after initial release.

    However the Quadro FX feature additional OpenGL support such as hardware clip planes make a significant difference and can be seen in SPECViewPerf benchmark results. For broadcast applications specialty Quadro FX products support GenLock and SDI outputs.

    Generally we've found the high-end Quadro FX products to have more onboard memory, to process larger dataset, than their GeForce cousins. Quadro FX products generally run at a more conservative core and/or memory clock rate compared to their GeForce products for extra margin of reliability.
    Listed below are quick specifications on the boards that we have reviewed:
    ProductQuadro FX 560Quadro FX 1500 Quadro FX 3500 Quadro FX 5500 Quadro FX 4600
    Price (retail)US$299US$699US$1,599US$2,999US$2,399
    Shader model3.03.03.03.04.0
    Core clock rate (MHz) 350325450650500
    Memory clock rate (MHz)1200 (DDR) 1250 (DDR) 1320 (DDR) 1010 (DDR) 1400 (DDR)
    Main memory128 MB GDDR3 256 MB GDDR3 256 MB GDDR3 1 BG GDDR2 768 MB GDDR3
    Memory interface width128-bit256-bit256-bit256-bit384-bit
    SLI support NoNoYesYesYes
    DVI output Two DVI-I (one dual-link)Two DVI-I (both dual-link) Two DVI-I (both dual-link) Two DVI-I (both dual-link) Two DVI-I (both dual-link)
    Stereoscopic outputNoNoYesYesYes
    SDI and G-Sync support NoNoNoYes, via options card Yes, via options card
    Component HD outputYes, via breakout box
    Yes, via breakout boxNoNoNo
    Max resolution

    3840 x 2400 @ 24Hz (dual link DVI)

    1920 x 1200 @ 60Hz (single-ink DVI)

    2048 x 1536 @ 75MHz (Analog)

    3840 x 2400 @ 24Hz (dual link DVI)

    1920 x 1200 @ 60Hz (single-ink DVI)

    2048 x 1536 @ 75MHz (Analog)
    3840 x 2400 @ 24Hz (dual link DVI)

    1920 x 1200 @ 60Hz (single-ink DVI)

    2048 x 1536 @ 75MHz (Analog)
    3840 x 2400 @ 24Hz (dual link DVI)

    1920 x 1200 @ 60Hz (single-ink DVI)

    2048 x 1536 @ 75MHz (Analog)
    3840 x 2400 @ 24Hz (dual link DVI)

    1920 x 1200 @ 60Hz (single-ink DVI)

    2048 x 1536 @ 75MHz (Analog)
    Bus interfacePCI Express 16xPCI Express 16xPCI Express 16xPCI Express 16xPCI Express 16x
    All the Quadro FX boards provide similar 2D capabilities, all have DVI-I outputs and the entry-level Quadro FX 560 even has one dual-link DVI output. How they differ is predominately 3D performance, amount of onboard memory, and specialized types of video outputs like SDI, GenLock, Component, and stereoscopic glasses support. The Quadro FX 4600 is the exception as its new unified-shader support opens the door to future applications.
    High-End Quadro FX 5500 and 4600

    The Quadro FX 5500 and 4600 DCC graphics boards gave the best performances in this review. Looking at the price tags, this comes as no surprise.

    The Quadro FX 5500, built around the G71GL GPU, was the strongest 3D workstation card from NVIDIA before the release of the Quadro FX 5600/4600. The card is a full-height PCI Express x16 measuring 241mm long. At the rear are two dual link DVI-I ports, and a stereoscopic connector for 3D shutter glasses. The top features an SLI connector and a 14-pin box connector for the GenLock daughter board.

    There is 1GB of onboard memory fed by a 256-bit bus. The board's power usage is 96 watts (quoted by NVIDIA) and a 6-pin Molex connector must be hooked up to the computer's power supply. The large fan assembly does it's job without too much noise generated even when the GPU is hottest under high load.

    The Quadro FX 4600 is the new addition to the Quadro FX workstation line. The only product in this review to offer Shader Model 4/DirectX 10 capabilities, the Quadro FX 4600's GPU is built upon the new G80GL GPU.

    The Quadro FX 4600 is a full-height PCI Express x16 card measuring 241mm long. The board's video output is two dual link DVI-I ports, and also has a stereoscopic connector for 3D shutter glasses. The top features an SLI connector and a 14-pin box connector for the GenLock daughter board. According to NVIDIA, the GenLock daughter board for the Quadro FX 5600/4600 will be different to the type for the Quadro FX 5500/4500.

    Onboard is 768MB of onboard memory fed by a 384-bit bus (actually six 64-bit buses). The board's power usage is 135 watts (quoted by NVIDIA), a 6-pin Molex connector must be hooked up to the computer's power supply. The Quadro FX 4600's large fan assembly is again, pretty quiet even when the GPU is under load.


      Next page
  • Mid-Level Quadro FX 3500 and 1500

    The middle of the range Quadro FX 3500 and 1500 look and perform similarly. The Quadro FX 3500 has SLI and stereoscopic connectors while the Quadro FX 1500 has component video output instead of stereoscopic output and no SLI connector.

    The Quadro FX 3500 is a full-height PCI Express x16 card measuring 210mm long. The board's video output two dual link DVI-I ports, and a stereoscopic connector for 3D shutter glasses. The top features an SLI connector. Onboard is 256MB of onboard memory. The board's power usage is 80 watts (quoted by NVIDIA), a 6-pin Molex connector must be hooked up to the computer's power supply.

    The Quadro FX 1500 is a full-height PCI Express x16 card measuring 210mm long, similar in appearance to the FX 3500. The board's video outputs are through two dual link DVI-I ports, and a component and S-video output Onboard is 256MB of onboard memory. The board's power usage is 65 watts (quoted by NVIDIA) and no auxiliary power is required.
    The Entry Level Quadro FX 560

    The Quadro FX 560 is the entry-level reviewed product, but it showed a respectable 3D performance. This probably tells me more about the aging benchmarks than anything else. The Quadro FX 560 is a full-height PCI Express x16 card measuring 188mm long. The board's video outputs include one single-channel DVI and one dual link DVI-I port. Component and S-video outputs are provided much like the Quadro FX 1500.

    Like on the FX 1500, the SLI connector was covered and unavailable. Onboard is 128MB of onboard memory. The board's power usage is 30 watts (quoted by NVIDIA) and no auxiliary power is required.
    Performance Benchmarks

    The G80GL GPU in the Quadro FX 4600 is designed to a unified-shader architecture. Inside the GPU are 96 running processors, each capable for executing both vertex and pixel shaders; compared with 24 pixel shader units and 8 vertex units in the Quadro FX 5500.

    The Quadro FX 4600's new unified-shader architecture will open new doors to future applications, as the GPU becomes more as commodity parallelizable co-processor.

    It's still early days, but interesting projects such as Folding@Home's GPU client do show that some compute operations can run very efficiently on GPU hardware. Another example is Apple's Core Image framework (used in Apple's Aperture application) that can utilize the GPU's pixel shaders to perform image-processing tasks.
    NVIDIA has recently announced it’s Compute Unified Device Architecture (CUDA) toolkit that runs on GeForce 8800-series, as well as Quadro FX 4600/5600 hardware that allow developers to create standard C language programs to execute on GPU hardware.

    At this time, the current G80GL Quadro FXs running CUDA perform IEEE-754 (with some deviations) single-precision arithmetic. According to the CUDA 0.8 release notes, future Quadro FX products will support double-precision math required for many science and engineering calculations. During our product briefing, NVIDIA say they are planning to offer a form of integrity-checking between the video card's onboard memory and GPU similar in fashion to ECC memory in workstations.

    While the hardware technology is certainly interesting and potentially useful, this article won't be covering the programmable shader capability of the new G80GL processor. The workstation benchmarks we're running utilize the graphics processor for traditional image synthesis.
    Overhauled Control Panel

    Over the years, NVIDIA's ForceWare control panel, have been functional but has never been exceptionally well laid out. They've had a go at overhauling it with a new Control Panel interface. The drivers we received for benchmarking is ForceWare 97.72 came with a new interface that follows more of a Windows XP-like wizard interface. Call me old-fashion but I prefer the more concise panel interface.

    While this new driver panel offers all the expected functionality (setting up resolution/color depth, multi-monitor, application profile etc), ForceWare 97.72 looked a little rough around the edges. For example "Video and Television" gave a corrupt overlay test image (video overlay didn't show up in screen capture), and there seemed to be a missing piece of text for the performance options.
    Software & Settings   Test Setup
    OSWindows XP 64-bit (SP1) Our workstation benchmark test used the following:

    • SPECViewPerf 9.0.3
    • SPECapc for 3ds max 8 (SP3)
    • SPECapc for Maya 6.5
    • CineBench 9.5

    Our test workstation is a BOXX Technologies 3DBOXX 8300, generously sent to us for the duration of our review.

    Notes:

    • All tests were performed according to SPEC run rules.
    • Texture compression were turned off.
    • nView profile were set to Maya and 3ds Max when running those applications respectively. SPECViewPerf and CineBench were set to default profile.
    • Vertical synchronization was turned off.

    A number of specific viewport driver settings were performed as per the SPEC run rules.

    The following dialog boxes capture the 3ds Max 8 configuration settings used: (click to enlarge)
    Video Driver ForceWare 91.36
    Quadro FX 5500, 3500, 1500, 560
    ForceWare 97.72
    Quadro FX 4600
    MaxTreme 8.00.03 16x
     
    DirectX9.0c 
    Monitor and screen resolution
    Sony G520 (analog VGA)
    1280x1024 pixels
    32-BPP
     

    Hardware
    Benchmark workstation3DBOXX 8300 
    CPUDual Intel Xeon X5355 @ 2.66GHz (Quad core) 
    Motherboard/chipsetTyan S3696 / Intel 5000X+ ESB2  
    System memory 4GB DDR2 667 FB-DIMM (ECC)  
    Hard Disk (Primary) Hitachi HDS721680PLA380 80GB 
    Previous page Next page
  • SPECViewPerf 9.0.3
    SPEC ViewPerf benchmark runs through scripted viewsets of popular CAD/CAM, digital content creation, and visualization applications. It's a synthetic benchmark because it characterizes the OpenGL performance and not the actual running application.

    SPECViewPerf aims to execute the exact OpenGL commands as would be sent by the application itself. SPECViewPerf 9.0.3's viewsets come from:

    • AutoDesk 3ds Max 3.1
    • Dassault Systemes CATIA V5R12
    • CEI EnSight
    • AutoDesk Lightscape Visualization System
    • AutoDesk Maya 6.5
    • PTC Pro/Engineer 2001
    • Dassault Systemes Solidworks 2004
    • UGS Teamcenter Visualization Mockup
    • UGS NX 3
    SPECViewperf 9.0.3
    (Higher is better)
     Quadro FX 560Quadro FX 1500 Quadro FX 3500 Quadro FX 5500 Quadro FX 4600
    3dsmax-0423.6927.6130.3932.8532.21
    catia-0236.5536.7538.940.0340.37
    ensight-0315.7619.5822.9626.1439.62
    light-0833.88
    34.0233.9934.0232.57
    maya-0266.61
    84.05106.6135.8165.7
    proe-0425.7329.3533.9238.4535.39
    sw-01 35.0739.6148.0659.2170.81
    tcvis-016.827.9210.5216.2623.2
    ugnx-018.249.5312.9419.1925.22
    All of the products were able to complete this benchmark without any issues. Even the entry-level Quadro FX 560 produced a reasonable result. The results show the Quadro FX 4600's strength in this particular benchmark, even exceeding the Quadro FX 5500 particularly in ensight-03 and tcvis-01 viewsets.
    SPECapc for 3ds Max 8 (SP3)
    SPECapc for 3ds Max 8 aims to measure the performance characteristics for a typical user's workflow in 3ds Max.
    The benchmark was run on all three available display drivers for 3ds Max 8: MAXtreme (Quadro accelerated OpenGL), standard OpenGL, and Direct3D.
    SPECapc for 3ds Max 8
    (Higher is better)
     Quadro FX 560Quadro FX 1500 Quadro FX 3500 Quadro FX 5500 Quadro FX 4600
    MAXTreme Driver     
    Wireframe Graphics3.013.773.493.804.13
    Mixed Wire/Shade GFX2.302.972.843.073.37
    Shaded Graphics
    2.603.673.413.134.22
    Graphics/Texturing
    /Lighting/Blending
    1.041.152.303.023.02
    Inverse Kinematics2.412.412.502.261.97
    Object Creation
    /Editing/Manipulation
    0.740.742.783.032.89
    Scene Creation Manipulation1.031.032.262.132.16
    CPU Render 9.369.329.639.759.56
    Overall Graphics
    1.712.072.813.023.37
          
    OpenGL Driver      
    Wireframe Graphics
    2.762.782.883.492.37
    Mixed Wire/Shade GFX1.491.481.491.771.34
    Shaded Graphics1.271.281.301.551.17
    Graphics/Texturing
    /Lighting/Blending
    0.640.751.190.751.17
    Inverse Kinematics
    1.851.831.841.831.43
    Object Creation
    /Editing/Manipulation
    1.341.431.961.501.78
    Scene Creation Manipulation1.451.451.891.451.66
    CPU Render
    9.629.519.559.559.48
    Overall Graphics1.121.191.441.321.32
          
    Direct3D Driver      
    Wireframe Graphics2.352.192.512.742.93
    Mixed Wire/Shade GFX2.302.182.572.813.23
    Shaded Graphics
    2.512.162.943.393.79
    Hardware Shaders4.676.186.797.086.75
    Graphics/Texturing
    /Lighting/Blending
    2.472.893.213.263.17
    Inverse Kinematics2.222.25
    2.222.232.27
    Object Creation
    /Editing/Manipulation
    2.532.672.762.912.68
    Scene Creation Manipulation2.092.142.142.152.21
    CPU Render9.689.599.729.639.57
    Overall Graphics2.432.422.863.073.21
    Overall Hardware Shaders4.676.186.797.086.75
    As SPECapc for 3ds Max 8 tests include CPU tests there aren’t as marked difference between the Quadro FXs. The overall results show that the more powerful Quadro 5500 and 4600 edging out the rest when using the Maxtreme and Direct3D tests.
    SPECapc Maya 6.5
    This benchmark aims to measure the typical user's workflow using Maya 6.5. Maya's graphics API used is OpenGL
    .
     
    SPECapc Maya 6.5
    (Higher is better)
     Quadro FX 560Quadro FX 1500 Quadro FX 3500 Quadro FX 5500 Quadro FX 4600
    Summary     

    GFX

    3.954.234.534.804.26
    I/O3.073.133.213.692.93

    CPU

    2.302.342.332.332.26

    Overall

    3.463.643.834.053.61
    Here the Quadro FX 5500 takes the top spot, and the scaling up of the different Quadro FXs seem much less pronounced. For some reason, the Quadro FX 4600 wasn't able to keep up with the Quadro FX 5500.
    CineBench 9.5
    This benchmark is based off MAXON's Cinema 4D application and uses OpenGL for display.
    .
     
    CineBench 9
    (Higher is better)
     Quadro FX 560Quadro FX 1500 Quadro FX 3500 Quadro FX 5500 Quadro FX 4600
    Summary     

    Shading (CINEMA 4D)

    513525515516526.00
    Shading (OpenGL Software Lighting)18571855186218651871.00

    Shading (OpenGL Hardware Lighting)

    58125926595259865915.00

    OpenGL Speedup

    11.3311.2911.5511.6111.24
    In this benchmark the scores hardly scaled with better performing video cards. I suspect the models are too simple for it to be an effective benchmark. I've left the results in for sake of completeness.
    Conclusion
    The Quadro FX 4600 and the Quadro FX 5500 proved to perform similarly. The Quadro FX 5500 has a full 1GB of RAM while the Quadro FX 4600 is built with 768MB; applications with huge data sets may favor the more expensive Quadro FX 5500. For many users the Quadro FX 4600 is likely the better and cheaper option.

    The Quadro 3500, 1500, 560 are positioned in their respective price points. As even the entry-level Quadro FX 560 performed well, the middle and entry boards they should well serve all but the most demanding users.

    Where SLI and stereoscopic glasses aren’t important, but two dual-link DVI, reasonable onboard memory, and high-quality component/s-video output are, then the Quadro FX 1500 would be my choice.

    If NVIDIA's roadmap is anything like in the past, as soon as the mid-range G80GL GPUs arrive, the middle and entry-level board will be replaced with G80GL-equivalents. In the short-term, graphics cards with unified-shaders, like the Quadro FX 5600 and 4600, offer the immediate benefit of performance improvements in current applications.

    They may also work as powerful parallel processors for future applications that rely on GPUs. It's going to be a fun ride.
    NVIDIA Quadro FX 4600
    Manufacturer: LeadTek Research
    Price: US$2,399

    Hits:
    - Very good performance matching or edging over last year's ultra top-end Quadro FX 5500
    - First workstation board to offer unified shader support. Expect the competition to close in soon though.
    - System ran without a hiccup across all tested applications.

    Misses:
    - Not cheap. However, if the Quadro FX 4600 is out of your price range, cheaper G80GL products are probably coming up.
    - The pre-release driver control panel interface looked like it needs more development time.

    Related links:

    NVIDIA


    {digg}
    Previous page More CGS Articles

blog comments powered by Disqus