[ Print Article! ]

AMD-ATI Radeon HD 2900 XT Performance Preview
May 14, 2007 Brandon Sandman Bell

Summary: After multiple delays, AMD's R600 GPU is finally here! Its got 320 stream processors for shading, a 512-bit memory interface, and over 700 million transistors. See how well this new GPU performs in comparison to GeForce 8800 GTX/GTS/Ultra in DX9 as well as our very first DX10 benchmarks, and don't forget the overclocking!


IntroductionPage:: ( 1 / 25 )

[image]

<% print_image("01"); %>


Much like ATI’s last next-gen introduction, R520, ATI is once again late to the game and playing catch up, only this time the stakes are higher, as this is ATI’s first major graphics launch under the combined AMD-ATI entity: make a mistake here, and it will affect the financials of the new AMD, a company which is already under a considerable amount of financial pressure. Of course, if the R600 launch is successful, AMD’s fortunes could swing in the other direction as the orders for DX10-compliant GPUs pour in from OEMs looking to refresh their mobile and desktop lineups, while enthusiasts pick up cards at the e-tail and retail levels. Needless to say this is a pretty big deal for AMD, and today’s opening salvo, the Radeon HD 2900 XT, has the specs on paper to really impress, but more on this later…

Let’s clear up what’s probably been the number one question we’ve received since the AMD-ATI buyout was announced: branding. Officially the latest Radeons will be branded as the ATI Radeon HD 2000 family of products from AMD. The Radeon HD 2000 line will start with the Radeon HD 2400 series, with prices starting below $99, while AMD’s mainstream solution, the Radeon HD 2600 series, will be priced from $99-$199. The flagship card will be the ATI Radeon HD 2900 XT from AMD, and will be priced at $399.

The Radeon HD 2900 XT should hit retail shelves starting today, while the other cards won’t arrive en masse until the end of June.

For the notebook market, AMD is also announcing new Mobility Radeon parts ranging from the Mobility Radeon HD 2300 for the value/entry-level segment, to the Mobility Radeon HD 2600 for performance notebooks. In between these two GPUs is the Mobility Radeon HD 2400 line, which will service the thin and light sector. According to AMD notebooks built around the Mobility Radeon HD 2300 should begin shipping in the May-June timeframe, but we won’t see laptops built with the other GPUs ship until sometime in July.

[image]
<% print_image("02"); %><% print_image("03"); %><% print_image("04"); %>

We’ll go over all the fine details of these parts later in this article, today we’re going to focus on AMD’s flagship part formerly known as R600, the Radeon HD 2900 XT. The specs for the Radeon HD 2900 XT have been leaked multiple times over the past few months, but we’ll provide them again just in case you missed it:

Unified Superscalar Shader Architecture

  • 320 stream processing units

    • Dynamic load balancing and resource allocation for vertex, geometry, and pixel shaders
    • Common instruction set and texture unit acces supported for all types of shaders
    • Dedicated branch execution units and texture address processors

  • 128-bit floating point precision for all operations
  • Command processor for reduced CPU overhead
  • Up to 80 texture fetches per clock cycle
  • Up to 128 textures per pixel
  • Fully associative multi-level texture cache design
  • DXTC and 3Dc+ texture compression
  • High resolution texture support (up to 8192x8192)
  • Lossless Z and stencil compression (up to 128:1)
  • Lossless color compression (up to 8:1)
  • 8 render targets (MRTs) with anti-aliasing support
  • Physics processing support

Full Support for DirectX 10.0
  • Shader Model 4.0
  • Geometry Shaders
  • Stream output
  • Integer and bitwise operations
  • Constant buffers
  • State objects
  • Texture arrays

    Dynamic Geometry Acceleration
  • High performance vertex cache
  • Programmable tessellation unit
  • Accelerated geometry shader path for geometry amplification
  • Memory read/write cache for improved stream output performance

    Anti-aliasing Features
  • Multi-sample anti-aliasing (up to 8 samples per pixel)
  • Up to 24x custom filter anti-aliasing (CFAA) for improved quality
  • Adaptive super-sampling and multisampling
  • Temporal anti-aliasing
  • Gamma correct
  • Super AA (CrossFire configurations only)
  • HDR+AA

    Texture Filtering Features
  • 2x/4x/8x/16x high quality adaptive anisotropic filtering modes (up to 128 taps per pixel)
  • 128-bit floating point HDR texture filtering
  • Bicubic filtering
  • sRGB filtering (gamma/degama)
  • Percentage Closer Filtering
  • Depth and stencil texture (DST) format support
  • Shared exponent HDR (RGBE 9:9:9:5) texture format support
    CrossFire Multi-GPU Technology
  • Scale up rendering performance and image quality with 2 or more GPUs
  • Integrated compositing engine
  • High performance dual-channel interconnect

    ATI Avivo HD Video and Display Platform

    • Two independent display controllers

      • Drive two displays simultaneously with independent resolutions, refresh rates, color controls, and video overlays for each display
      • Full 30-bit display processing
      • Programmable piecewise linear gamma correction, color correction, and color space conversion
      • Spatial/temporal dithering provides 30-bit color quality on 24-bit and 18-bit displays
      • High quality pre- and post-scaling engines, with underscane support for all display outputs
      • Content-adaptive de-flicker filtering for interlaced displays
      • Fast, glitch-free mode switching
      • Hardware cursor

    • Two integrated dual-link DVI display outputs

      • Each supports 12-, 24-, and 3-bit digital displays at all resolutions up to 1920x1200 (single-link DVI) or 2560x1600 (dual-link DVI)
      • Each includes a dual-link HDCP encoder with on-chip key storage for high resolution playback of protected content

    • Two integrated 400MHz 32-bit RAMDACs

      • Each supports analog displays connected by VGA at all resolutions up to 2048x1536

    • HDMI output support

      • Supports all display resolutions up to 1920x1080
      • Integrated HD audio controller with multi-channel (5.1) AC3 support, enabling a plug-and-play cable-less audio solution

    • Integrated Xileon HDTV encoder

      • Provides high-quality analog TV output (component/S-Video/composite)
      • Supports SDTV and HDTV resolutions
      • Underscan and overscan compensation

    • HD Decode Acceleration for H.264/AVC, VC-1, DivX and MPEG-2 video formats

      • Flawless DVD, HD-DVD, and Blu-ray playback
      • Motion compensation and IDCT (inverse discrete cosine transformation)

    • HD Video Processing

      • Advanced vector adaptive per-pixel de-interlacing
      • De-blocking and noise reduction filtering
      • Edge enhancement
      • Inverse telecine (2:2 and 3:2 pull-down correction
      • Bad edit correction
      • High fidelity gamma correction, color correction, color space conversion, and scaling

    • MPEG-2, MPEG-4, DivX, WMV9, VC-1, and H.264/AVC encoding and transcoding
    • Seamless integration of pixel shaders with video in real time
    • VGA mode support on all display outputs

    512-bit, 8-channel GDDR3/4 Memory Interface

    • Ring-Bus Memory controller

      • Fully distributed design with 1024-bit internal ring bus for memory reads and writes
      • Optimized for high performance HDR (high dynamic range) rendering at high display resolutions


    700 Million Transistors on 80-nm HS fabrication process
    OpenGL 2.0 Support
    PCI Express x16 bus interface

    Notes

    On paper AMD’s Radeon HD 2900 XT certainly looks impressive. It boasts over 320 stream processing units and a 512-bit memory interface, that’s twice as wide as ATI’s previous high-end offering, the Radeon X1950 XTX.

    R600’s stream processors are arranged as a 5-way design, with its memory interface consisting of eight 64-bit memory controllers; in comparison the Radeon X1950 XTX relied on eight 32-bit memory controllers for its 256-bit memory interface. The chip is built on a special high-speed derivative of TSMC’s 80-nm manufacturing process (normally 80-nm is reserved for value and mainstream parts) and consists of over 700 million transistors. That’s nearly twice the number of transistors as Radeon X1950 XTX, which incorporated 384 million transistors, and 19 million more transistors than the G80 GPU inside NVIDIA’s GeForce 8800.

    Another feature that’s been discussed extensively is audio. Rumors have swirled that R600 boasts an integrated 5.1 audio controller, leading many to believe that R600 would handle audio-processing duties, but we can report that ATI’s audio solution is merely a pass-through solution using the PCI Express interface to connect to your existing audio card. This is useful for home theater PC (HTPC) users who would like to connect their HDTV to their PC over HDMI. Previously you had to run a separate cable from your graphics card to your sound card in order to pass audio to your HDTV via HDMI. With AMD’s latest Radeons, this cable is no longer necessary.

    Looking over the specs, you’ll also no doubt see ATI’s new 24x custom filter AA mode.




    New unified architecturePage:: ( 2 / 25 )

    High-end Specifications Comparison
    Radeon HD 2900 XTGeForce 8800 GTXGeForce 8800 GTS 640MB
    # of Transistors700M681M681M
    Core Clock Speed (MHz)742575500
    Stream Processor Clock Speed (MHz)74013501200
    # of Stream Processors32012896
    Texture Units163224
    ROPs162420
    Texture fill-rate (Gigatexels/sec)11.918.412.0
    Memory Clock (MHz)165018001600
    Memory Interface512-bit384-bit320-bit
    Memory Bandwidth (GB/sec)105.686.464
    Memory Size512MB768MB640MB



    While R600 is technically a new design, the chip leverages many technologies already found in ATI’s last two GPUs: R520/580 from the Radeon X1000 series and the Xbox 360’s Xenos GPU. Looking over a block diagram of R600, you’ll immediately recognize many units from R5xx. For instance, to improve dynamic branching, ATI continues to break down the processing workload into a large number of small threads. These threads are then managed by the ultra-threading dispatch processor. Meanwhile from Xbox 360 ATI leverages their unified shading architecture.

    For R600, ATI merely builds on this design, adding more powerful superscalar shader processors, a tweaked ultra-threading dispatch processor, the addition of a new tessellation unit (actually borrowed from Xenos) full DirectX 10 support, and a more robust, 512-bit memory interface. All this adds up to a GPU that’s been designed for the next generation in HD gaming, delivering very high levels of performance even when gaming at mega resolutions such as 2048x1536 and 2560x1600 with HDR+AA.

    At the heart of all this is ATI’s 2nd-generation unified shader architecture. Consisting of 320 distinct, independent stream processing units, it’s quite impressive. Like NVIDIA’s G80 GPU, ATI has incorporated a scalar architecture for R600’s shading processors (the stream processing units). Only in ATI’s case, R600 can issue many more independent instructions in each shader processor due to its superscalar design. The Radeon HD 2900 can issue up to five scalar multiply-add (MAD) operations and one branch instruction to each shader processor per clock cycle. In comparison, each stream processor in G80 can dual-issue a MAD and MUL instruction.

    [image]

    <% print_image("05"); %>

    In the block diagram above the stream processors are depicted as yellow squares in the center of the GPU. Attached to each group of five stream processors is a dedicated branch execution unit (the purple square) and general purpose registers which can be used to store input data, temporary values, and output data. Here’s a group of stream processors up close:

    With more shading units onboard, it’s important to keep these shader processors fed with data. This is where the ultra-threading dispatch processor comes in. The ultra-threading dispatch processor acts as a traffic cop, it’s a central dispatch unit that is responsible for tracking and distributing thousands of threads simultaneously across the Radeon HD 2900’s shader processors. ATI won’t provide a specific max number of threads, but in comparison, R520/580 was limited to 512 threads.

    [image]
    <% print_image("06"); %>

    The above diagram breaks down the ultra-threading dispatch processor. As you can see, ATI has provided separate command queues for each shader type: pixel, vertex, or geometry data. From there arbiter units determine which threads will be processed first, based on a variety of parameters. ATI provides two arbiter units per SIMD array, with dedicated arbiter units for texture and vertex fetches, allowing them to be scheduled independently of math operations. Threads that are already executing can be bumped at any time if a higher priority thread is pulled from the command queues. The temporary data is saved so the thread can be resumed later. If a thread is forced to wait for data, it is suspended and a new thread begins executing immediately. The suspended threads remain in the command queue until their requested data arrives. According to ATI, hundreds of threads can be queued up to make sure the SIMD arrays are never sitting idle.



    Architecture (cont’d)Page:: ( 3 / 25 )

    Tesselation unit

    At the top of the 2900 XT block diagram is the setup engine, which is responsible for preparing data for processing by the stream processing units. The setup engine consists of five core components:

  • Scan converter/rasterizer and interpolators (for pixel shaders)
  • Geometry assembler (for geometry shaders)
  • Vertex assembler (for vertex shaders)
  • Programmable tessellation unit

    The Radeon HD 2000 series is the first GPU to feature a dedicated hardware unit for tessellation. With tessellation, developers can create complex objects without requiring a large amount of triangles. For example, the GPU could be used to dynamically render features such as the topology of a mountain range. With adaptive tessellation, the developer can selectively put more detail where needed (via more patches), and less patches where high detail isn't as important. For example, more distant objects can have less detail than objects that are closer to the viewer. This allows highly detailed scenes to be rendered with much fewer polygons than would otherwise be necessary. ATI’s tessellation unit supports a wide variety of higher order surfaces, including Bezier patches, N-Patches, and NURBS.

    Another benefit of ATI’s tessellation unit is faster displacement mapping and reduced memory bandwidth usage. This provides the greatest benefit to lower-end cards like the Radeon HD 2400 and 2600, which don’t have the horsepower of the Radeon 2900 XT when it comes to triangle processing rate and memory bandwidth.

    To ease development for game developers, ATI has created a software library developers can use to code for their tessellation unit. For DX10 hardware that lacks a dedicated tessellation unit (read: NVIDIA’s GeForce 8 line), ATI’s tools would utilize the geometry shader to provide similar functionality. For DX9 cards, which obviously lack support for geometry shaders, the CPU would be used instead.

    The biggest challenge standing in the way of the tessellation unit will likely be game developer support. Here AMD argues that future APIs will include support for hardware-based tessellation, in particular citing a GDC 2007 presentation from Chas Boyd titled “The Future of DirectX”. AMD also argues that the tessellation unit found inside the Radeon HD 2000 series is the exact same unit game developers are already using today inside Xbox 360. In fact, one popular Xbox 360 game, Viva Piñata, utilizes the tessellation unit extensively.

    AMD isn’t ready to disclose which upcoming PC games will take advantage of the tessellation unit, but when we asked for a rough figure we were told to expect a handful of games by the end of this year.


    Video

    With HD-DVD and Blu-ray playback on the PC becoming increasingly popular, handling H.264 and VC-1 decoding on the GPU has become a higher priority for both AMD and NVIDIA. To tackle this problem, both companies have integrated dedicated resources on the GPU for handling video decoding. In the case of AMD, their video hardware is known as UVD, which is short for universal video decoder. AMD’s UVD is capable of handling the entire decode process and supports full 40Mbps bit-rates. The following graphs summarize the benefits of UVD:

    [image]

    <% print_image("07"); %><% print_image("08"); %>

    As you can see, with UVD and Avivo HD, the Radeon HD 2000 series is capable of handling both the bitstream processing and frequency transform stages for both VC-1 and H.264 playback, whereas the VP2 processor in NVIDIA’s GeForce 8600/8500/8400 is only capable of handling this for high definition movies using the H.264/AVC Codec; when using VC-1, this is handled by the CPU on the GeForce cards.

    With UVD handling all decode stages for high definition playback, CPU utilization for high definition playback is just as low as it was for standard definition playback. This is important for those of you with slower CPUs and/or those on notebooks. Without the GPU handling these tasks, CPU utilization for a high bit-rate HD-DVD or Blu-ray disc could be over 70%, even on a fast Core 2 Duo CPU.

    In addition, ATI has added custom logic to the Radeon HD 2400 and 2600 for video post-processing capable of performing de-interlacing, up/down scaling, and color correction.

    All Radeon HD 2000 series GPUs are HDCP-compliant. In fact, the Radeon HD series is the first line of GPUs to integrate the HDCP encryption keys directly into the ASIC, no external CryptoROM chip is needed. As a result, we’d estimate that all Radeon HD 2000 series cards should support HDCP, regardless of manufacturer. Previously this was often left in the hands of board manufacturers, particularly on lower end cards where board partners didn’t want the added expense of adding HDCP support. All Radeon HD 2000 series cards are also capable of supporting HDCP over dual-link DVI.

    [image]
    <% print_image("09"); %><% print_image("10"); %>

    When it comes to audio, as we mentioned before, the Radeon HD 2000 series is capable of transmitting both audio and video over HDMI. No external audio connection is required. AMD and their board partners will be bundling HDMI adapters with their cards which are capable of delivering full audio and video via the standard DVI output on the back of the Radeon card.



    The Radeon HD 2900 XT CardPage:: ( 4 / 25 )


    There were rumors swirling of multiple Radeon HD 2900 SKUs in development: a Radeon HD 2900 XT like the card you see today, a 9.5” Radeon HD 2900 XTX card intended for the retail market, and a OEM variant of the XTX which featured a 11.5” PCB and slightly different cooler. Ultimately though only one SKU is being released today – the Radeon HD 2900 XT.

    [image]

    <% print_image("11"); %><% print_image("12"); %><% print_image("13"); %>

    AMD won’t confirm if they have any plans to eventually release an XTX model; assuming an XTX model was in the works for release early this summer we prodded multiple representatives from ATI on this issue and we consistently got a negative response, suggesting that if AMD does plan to eventually reveal a Radeon HD 2900 XTX SKU with faster clocks it isn’t due for release in the immediate future. Our guess is we won’t see an ultra high-end $500+ SKU until R600 transitions to TSMC’s smaller, less power-hungry 65-nm manufacturing process.

    Power delivery

    Another aspect that’s been discussed repeatedly on numerous forums across the web relates to the Radeon HD 2900 XT’s power draw. These rumors were fueled largely because of the card’s unique 8-pin power connector. Located in the upper right corner you’ll immediately spot the 6-pin PCI Express power connector that’s become standard fare on most high-end graphics cards, however sitting alongside it is a larger 8-pin power connector.

    [image]

    <% print_image("14"); %>

    This second, 8-pin power connector is required in order for the card to operate, but fortunately it’s backward-compatible with today’s existing power supplies, many of which only have 6-pin PCIe connectors. You can see how it’s been designed to operate in the following image:

    [image]
    <% print_image("15"); %>

    Now you’re probably wondering why AMD integrated an 8-pin power connector on the Radeon HD 2900 XT if the card can run off all of today’s existing power supplies, many of which only have 6-pin PCIe connectors. The 8-pin power connector is only required for overclocking.

    You see, the Radeon HD 2900 XT’s maximum TDP at stock speeds is 215 watts (in comparison, the max TDP for GeForce 8800 GTX is 185 watts). Each auxiliary 6-pin PCIe power connector can run up to 75W, and the PCIe interface itself maxes out at 75W. With all three connections, that’s 225 watts – just enough power to get by at stock speeds. That’s not enough power for someone who may be overclocking their card however. This is where the 8-pin power connector comes in.

    The 8-pin power connector alone can carry up to 150 watts. When combined with the 75W from the PCIe interface and the 75 watts from the 6-pin PCIe power connector, up to 300W can be fed to the Radeon HD 2900 XT. We conducted power consumption tests later in this article that you’ll want to check out, but to make a long story short, you won’t need a PSU with an 8-pin power connector unless you want to overclock the graphics card.

    In order to overclock the card, you must plug in the 8-pin power connector and the 6-pin power connector, otherwise the driver automatically disables all overclocking.

    In terms of power supply guidelines, AMD recommends a minimum of a 550W power supply for single card Radeon HD 2900 XT configurations, and a 750W PSU if you plan on running CrossFire.



    Board analysis (cont’d)Page:: ( 5 / 25 )

    The cooling

    Besides the Radeon HD 2900 XT’s unusual power requirements, the other feature enthusiasts have been talking about is the Radeon 2900 XT’s new cooling unit. As you can see in the photos, it’s nothing like the cooler ATI employed previously on the Radeon X1950 XTX, while the card does rely on a ducted cooling design, and the board’s fan does blow hot air outside your system case, it’s a much larger cooler.

    [image]

    <% print_image("16"); %><% print_image("17"); %><% print_image("18"); %>

    AMD now uses two heat pipes (versus one on the X1950 XTX), while a longer heatsink is responsible for keeping the heat pipes cool.

    In its current form, AMD’s cooler is quite loud, but we wouldn’t be surprised if this is tweaked with a subsequent driver release. Part of the reason why we think this could be adjusted in the future is because ATI has done this before with the Radeon X1800 XT/XL and X1900 XT/XTX. Initially the fans on these cards ran ridiculously loud, but eventually ATI tweaked them so they ran at more reasonable levels. Another reason why we say this is because the 2900 XT fan currently only runs at two basic settings (a barely audible low setting and a much louder high setting) that leaves a lots of room for AMD to tweak the RPMs if they desire to do so. As you’ll see in our temperature testing, the GPU doesn’t get that hot under full load in large part because the fan spins at such high RPMs.

    Hopefully with a little bit of tweaking on AMD’s part the fan can be adjusted to run quieter without affecting temperature too drastically. Personally, we think the reason why AMD decided to run the fan speeds so high is to keep their power consumption levels down. During AMD’s Editor’s Day event it was revealed that the Radeon 2900 XT’s power consumption depended greatly on its temperature. Referring back to our notes, it was revealed that “the ambient temp the GPU runs at will very much affect the amount of power the board consumes”. Quite simply, the hotter the board runs, the more power it will draw.

    This all goes back to the way Radeon 2900 XT GPUs are binned. Since there is no XTX SKU, all R600 chips are being used in the Radoen HD 2900 XT. Normally the chips that consume the least power and generate the least heat would go into the XTX SKU, but since ATI was never able to get R600’s power consumption under control, the XTX SKU no longer exists and all R600 chips are going into XTs. As a result, we were warned that there could be a lot of variance in how much power different Radeon HD 2900 XT cards consumed, and to keep temps down in order to keep power consumption in check. As it stands now ATI runs the fan on the Radeon HD 2900 XT at higher RPMs in order to keep temps, and thus power consumption down to manageable levels. As a result though, the board generates considerably more noise than the GeForce 8800 GTS/GTX. In fact, under load two GeForce 8800 cards generate less noise than one Radeon HD 2900 XT.


    VIVO and Valve returns to offer another game bundle

    One feature that ATI, and now AMD continues to provide in their high-end cards is VIVO (video-in/video-out). The Radeon HD 2900 XT is no exception: all cards will support VIVO. In fact, ATI has replaced the venerable Rage Theater chip used previously to provide this functionality with their newer Theater 200 chip.

    Another feature that will be universal to all Radeon HD 2900 XT cards is the game bundle. Like the Radeon 9800 XT launch, ATI has once again partnered with Valve to deliver an excellent game bundle. In this case, all cards will ship with the Black Box Edition of Half-Life 2, which not only includes the latest chapter in Valve’s Half-Life 2 storyline, Half-Life 2 Episode Two, but also includes Portal and Team Fortress 2.

    [image]

    <% print_image("19"); %><% print_image("20"); %>

    Sapphire’s Radeon HD 2900 XT card

    Sapphire was the first AMD board partner to get a final, shipping retail card in our hands, and will likely be first to market with cards at e-tail right alongside AMD’s own card.

    As you can see, Sapphire’s Radeon HD 2900 XT card is based entirely on the reference board design for the Radeon HD 2900 XT, literally the only difference between the two cards is the sticker on the fan. Sapphire ships their Radeon HD 2900 XT card with everything you’d need to get up and running, including the HDMI adapter, CrossFire cable, VIVO cable, component video cable, and two DVI adapters. Meanwhile on the software side Sapphire includes DVD playback software, the Half-Life 2 Black Box voucher and a copy of the Pro version of 3DMark 06.

    But Sapphire isn’t sticking to just one SKU for the Radeon HD 2900 XT. For enthusiasts looking for more performance, Sapphire will also be offering a Toxic Edition of the card. The Sapphire Radeon HD 2900 XT Toxic will ship with more memory than the standard Radeon HD 2900 XT – 1GB of GDDR3, and both the graphics core and memory will be overclocked out-of-the-box. Sapphire’s currently projecting an OC of 15% for the graphics core, and 50-70MHz for the board’s memory.

    [image]

    <% print_image("21"); %>

    If that weren’t enough, the Toxic will also be liquid cooled. Shipping with the card is an external water cooling unit which will require two empty 5.25” drive bays. Keeping everything cool will be a large fan outfitted with blue LEDs that will spin anywhere from 1600-2400 RPMs. The cooling system is completely modular, meaning you can add an additional Radeon HD 2900 XT Toxic card for CrossFire, or even a 3rd-party CPU waterblock for cooling your processor.

    The Toxic card is expected to ship sometime in June, and we’ll definitely be keeping our eyes out for it.

    A new driver arrives

    Late last week AMD submitted a pre-release alpha driver which delivered their new 12x and 24x AA modes, as well as providing performance enhancements in Oblivion and other apps with HDR lighting. According to AMD:


    “The driver team has been working on improving adaptive AA performance over the existing press driver with a new intelligent algorithm that increases FPS while applying similar image quality to that of the first press driver. In some cases, such as Oblivion, we're seeing performance that is several times faster than what was seen before using the new adaptive AA algorithm. We've also done some new optimizations for HDR applications that in general result in a 5-30% increase in performance. This is an absolutely incredible performance increase within just a few short weeks of the initial driver build, and a testament to the programmable capabilities of the architecture.”


    Based on this, we decided to re-run our tests with the new driver. In the subsequent graphs you’ll see the Radeon HD 2900 XT running with this driver labeled as “alphadriver”. Unfortunately by conducting these tests we weren’t able to take a closer look at the image quality and performance of AMD’s new AA modes, but we will take a look at that shortly in a subsequent article.



    Radeon HD 2600/2400 SeriesPage:: ( 6 / 25 )

    ATI Radeon HD 2000 Series
    ATI Radeon HD 2400ATI Radeon HD 2600ATI Radeon HD 2900 XT
    Stream Processing Units40120320
    Clock Speed525-700MHz600-800MHz740MHz
    Math Processing Rate (Multiply-Add)42-56 GigaFLOPS144-192 GigaFLOPS475 GigaFLOPS
    Pixel Processing Rate4.2-5.6 Gigapixels/sec14.4-19.2 Gigapixels/sec47.5 Gigapixels/sec
    Triangle Processing Rate262-350M triangles/sec600-800M triangles/sec742M triangles/sec
    Texture Units4816
    Render Back-Ends4416
    Memory Frame Buffer Size256MB GDDR3, 128MB/256MB DDR2256MB GDDR4, 256MB GDDR3, 256MB DDR2512MB GDDR3
    Memory Interface Width64-bit128-bit512-bit
    Memory Clock400-800MHz400-1100MHz825MHz
    Memory Bandwidth (GB/sec)6.4-12.8 GB/sec12.8-35.2 GB/sec106 GB/sec
    Transistors180M390M700M
    Technology65G+65G+80HS
    Price<$99$99-$199$399
    OutputsSVGA+DDVI+VO (HDMI adapter)D+DL+DVI (HDMI adaptor)D+DL+DVI w/HDCP (HDMI adaptor)



    Launching alongside the Radeon HD 2900 XT today are two additional product lines, the Radeon HD 2600 and Radeon HD 2400. As we mentioned earlier, while Radeon HD 2900 XT cards are shipping today, the 2600 and 2400 series won’t be available until the end of next month.

    Both GPUs are built largely on the same architecture found in R600, only features have been removed to cut costs. As you can see in the chart above, these mainstream and value GPUs are endowed with fewer functional units and narrower memory interfaces. The following block diagrams summarize the changes nicely:

    [image]

    <% print_image("22"); %><% print_image("23"); %>

    Radeon HD 2600 Series

    Taking on the GeForce 8600 line is AMD’s Radeon HD 2600 series. The Radeon HD 2600 is built on AMD’s RV630 GPU and thanks to its fewer shaders and 65-nm manufacturing process, it consumes significantly less power than R600. According to AMD, TDP is just 45-watts for RV630. As a result, RV630 doesn’t need an external PCIe power connection in order to operate.

    [image]

    <% print_image("24"); %><% print_image("25"); %>

    Two SKUs are currently planned for the Radeon HD 2600 series: the Radeon HD 2600 Pro and the Radeon HD 2600 XT. While clock speeds aren’t final yet, we’ve been told they’ll be in the neighborhood of 600-800MHz.

    [image]
    <% print_image("26"); %><% print_image("27"); %>

    The Radeon HD 2600 series will support AMD’s built-in CrossFire technology, allowing enthusiasts to pair two cards together for double the performance. As you can see in the board shots, both RV630 boards will ship with single-slot coolers and include an HDMI adaptor, just like the 2900 XT. Unlike the Radeon HD 2900 XT however, these cards won’t ship with the Half-Life 2 Black Box.

    Radeon HD 2400 Series

    The Radeon HD 2400 series is AMD’s entry-level DX10 solution aimed at the value market. AMD’s been pretty adamant that board prices will definitely be below $99. This low pricing, and AMD’s use of the 65-nm process, could make these cards ideal for use in a HTPC, particularly if you’re looking for a low-cost video card for handling HD video. Like the Radeon HD 2600 series, two SKUs are planned: the Radeon HD 2400 Pro, and the Radeon HD 2400 XT.

    [image]

    <% print_image("28"); %><% print_image("29"); %>

    Once again clock speeds haven’t been finalized and we won’t know more about the performance of these cards for another month, but considering the compromises that have been made to the graphics core and memory subsystem, you’ll probably want to step up to a Radeon HD 2600 series or better for serious gaming. These cards would likely make a nice upgrade over an integrated graphics solution however.

    [image]
    <% print_image("30"); %><% print_image("31"); %>

    The Radeon HD 2400 series is based on AMD’s RV615 GPU and boasts a TDP of just 25 watts.




    Test SystemsPage:: ( 7 / 25 )

    System Setup


    Intel Core 2 Extreme X6800

    EVGA nForce 680i SLI motherboard
    ASUS P5W DH Deluxe (for Radeon HD 2900 XT cards)

    2GB Corsair TWIN2X2048-6400C4

    ATI Radeon HD 2900 XT
    Sapphire Radeon HD 2900 XT
    HD2900_xp_8-37-4-070419a-046506E
    HD2900_vista_8-37-4-070419a-046507E
    ATI Radeon X1950 XTX
    Catalyst 7.4

    NVIDIA GeForce 7900 GTX
    ForceWare 93.71

    BFG GeForce 8800 GTS 640MB OC2
    NVIDIA GeForce 8800 GTS 640MB
    NVIDIA GeForce 8800 GTX 768MB
    NVIDIA GeForce 8800 Ultra 768MB
    EVGA e-GeForce 8800 GTS KO ACS3 640MB
    ForceWare 158.22

    300GB Western Digital Caviar SE

    Windows XP Professional SP2


    Benchmarks

    Call of Juarez DX10 Beta Benchmark
    Company of Heroes 1.3
    F.E.A.R. 1.08
    Elder Scrolls IV: Oblivion
    Battlefield 2142 1.10
    Call of Duty 2 1.3
    Half-Life 2 Lost Coast
    Quake 4 1.3
    STALKER 1.001
    Far Cry
    3DMark 06

    Notes

    To save time during testing, we used two different factory overclocked GeForce 8800 GTS 640MB boards: EVGA’s e-GeForce 8800 GTS KO ACS3, and the BFG GeForce 8800 GTS OC2 (two OC2 BFG cards were used exclusively for SLI testing). Both boards are factory overclocked to the same speeds, so performance between the two should be similar. Even though the Radeon HD 2900 XT is priced to compete with the GeForce 8800 GTS 640MB line, we still went ahead and included the GeForce 8800 GTX/Ultra to see how NVIDIA’s flagship cards stack up against AMD’s latest and greatest.



    3DMark 06Page:: ( 8 / 25 )

    3DMark 06 – Direct3D







    Battlefield 2142Page:: ( 9 / 25 )

    Battlefield 2 – Direct3D





    Notes

    Battlefield 2142 is based on the same game engine used for Battlefield 2, and therefore doesn’t use shaders extensively like some of today’s newer games. As a result, the extra texturing units found in the GeForce 8800s help them pull away from the Radeon HD 2900 XT cards in this game. In fact, the performance of the Radeon HD 2900 XT closely mirrors that of the Radeon X1950 XTX, which has the same number of texturing units and ROPs as the Radeon HD 2900 XT.

    What’s really holding the Radeon HD 2900 XT’s performance back in this game however is probably the graphics driver.




    F.E.A.R. PerformancePage:: ( 10 / 25 )

    F.E.A.R. – Direct3D






    FEAR Performance 1600x1200x32
    CardMin FPSMax FPS
    GeForce 8800 Ultra49235
    GeForce 8800 GTX44211
    GeForce 8800 GTS 640MB32140
    GeForce 8800 GTS OC236158
    Radeon HD 2900 XT22139
    Radeon HD 2900 XT CrossFire25315
    GeForce 8800 GTS OC2 SLI56369
    Radeon HD 2900 XT Alphadriver23137
    Sapphire Radeon HD 2900 XT23139


    Notes

    The Radeon HD 2900 XT’s performance improves dramatically in comparison to the GeForce 8800 cards in FEAR, which is one of the more shader-intensive games that we test with. As you can see, the 2900 XT still falls behind the factory overclocked GeForce 8800 GTS card at lower screen resolutions, but once the resolution is increased and memory bandwidth becomes more of a factor, the Radeon 2900 XT is able to pull ahead thanks to its 512-bit memory interface. Keep in mind that this doesn’t occur until 2560x1600 though, which is a higher resolution than most users will game at.




    Oblivion Mountains HDRPage:: ( 11 / 25 )

    Oblivion – Direct3D






    Oblivion Performance 1600x1200x32
    CardMin FPSMax FPS
    GeForce 8800 Ultra87146
    GeForce 8800 GTX76124
    GeForce 8800 GTS 640MB61100
    GeForce 8800 GTS OC266102
    Radeon X1950 XTX3555
    Radeon HD 2900 XT5079
    Radeon HD 2900 XT CrossFire82118
    Radeon HD 2900 XT Alphadriver5078
    Sapphire Radeon HD 2900 XT5079


    Notes

    Oblivion is another shader-heavy title that we test with, although both the factory overclocked and stock GeForce 8800 GTS cards are able to remain ahead by a comfortable margin of 20%+, even at 2560x1600. In fact, the Radeon X1950 XTX only trails the Radeon HD 2900 XT cards by 18-25% in this benchmark, that’s certainly not what you’d expect from a next-gen product.




    Oblivion Foliage HDRPage:: ( 12 / 25 )

    Oblivion – Direct3D






    Oblivion Performance 1600x1200x32
    CardMin FPSMax FPS
    GeForce 8800 Ultra5476
    GeForce 8800 GTX5172
    GeForce 8800 GTS 640MB3754
    GeForce 8800 GTS OC24358
    Radeon X1950 XTX2735
    Radeon HD 2900 XT3954
    Radeon HD 2900 XT CrossFire7193
    Radeon HD 2900 XT Alphadriver3853
    Sapphire Radeon HD 2900 XT3855


    Notes

    Under our more demanding foliage testing, the Radeon HD 2900 XT is able to pull closer to the GeForce 8800 GTS 640MB, actually overtaking the stock card by 1600x1200, but it’s not quite able to outpace the factory overclocked e-GeForce 8800 GTS 640MB ACS3. We have a strong suspicion that AMD’s got a lot more driver work to do when it comes to optimizing performance for this game though, as the results we’re seeing here just don’t correspond to what we expected.



    Call of Duty 2Page:: ( 13 / 25 )

    Call of Duty 2 – Direct3D






    Notes

    Our results with Call of Duty 2 are probably the most shocking of all – the Radeon X1950 XTX actually outran the Radeon HD 2900 XT! We confirmed with AMD that a driver issue is hampering performance, they’re seeing similar results in their own internal testing “We’ve apparently got a driver issue that sees the X1950 XTX beat the 2900 XT. Andy and the driver folks are still ironing out the cause though.”

    Needless to say, these results shouldn’t be taken too seriously until AMD is able to resolve what’s causing the performance problems.



    HL2 Lost Coast HDRPage:: ( 14 / 25 )

    Half-Life 2 Lost Coast – Direct3D








    Company of HeroesPage:: ( 15 / 25 )

    Company of Heroes – Direct3D






    Notes

    The Radeon HD 2900 XT performs extremely well in Company of Heroes. Not only does it outrun the GeForce 8800 GTS 640MB in this benchmark, it eventually manages to catch up to the GeForce 8800 GTX and then overtake it at 2560x1600. We’ll be curious to see how the Radeon HD 2900 XT performs once the DX10 patch for Company of Heroes patch is ultimately released.



    Quake 4Page:: ( 16 / 25 )

    Quake 4 – OpenGL






    Notes

    While ATI’s OpenGL performance has generally taken a backseat to NVIDIA’s in the past, this time around the Radeon HD 2900 XT is able to pull ahead pretty comfortably from the stock GeForce 8800 GTS and factory overclocked EVGA card, especially at 2560x1600. You can really see the Radeon HD 2900 XT’s memory bandwidth in action here, as it even begins to reel in the GeForce 8800 GTX, but doesn’t quite catch it by 2560x1600.



    STALKERPage:: ( 17 / 25 )

    STALKER – Direct3D









    Far CryPage:: ( 18 / 25 )

    Far Cry – Direct3D






    Far Cry Performance 1600x1200x32
    CardMin FPSMax FPS
    GeForce 8800 Ultra89.7168.4
    GeForce 8800 GTX81.5153.8
    GeForce 8800 GTS 640MB61.5108.3
    GeForce 8800 GTS OC268.5122.3
    Radeon X1950 XTX50.975.8
    Radeon HD 2900 XT72.5128.2
    Radeon HD 2900 XT CrossFire78.7214.7
    GeForce 8800 GTS OC2 SLI88.4215.2
    Radeon HD 2900 XT Alphadriver63.5126.9




    DX10 Gaming: Call of JuarezPage:: ( 19 / 25 )

    Call of Juarez – Direct3D10






    Notes

    One of the first games that will be patched to support DX10 graphics is Call of Juarez. Reviewers were given the opportunity to test a very early beta build of the game, the results of which you see above. In testing with the early build of the game, the Radeon HD cards are able to pull ahead of the factory overclocked GeForce 8800 GTS card by about 6-8%, but still trail the GeForce 8800 GTX and Ultra. However this early build of the game has an application bug that prevents GeForce 8800 cards from rendering the game with AA. We were able to get our hands on a newer build of the game and obtained the following results:






    Under the new build of the game, the factory overclocked GeForce 8800 GTS card is able to squeeze ahead of the Radeon HD 2900 XT under both 0xAA and 4xAA. In fact, in both cases the GeForce 8800 GTS card is able to pull away from the Radeon HD 2900 XT as resolution is increased. This is surprising considering the Radeon card’s memory bandwidth advantage.

    We should hopefully see more DX10 content very soon. The DX10 patch for Company of Heroes should be released later this month, while a DX10 version of the hit Xbox 360 shooter Lost Planet is expected to be released later this summer.



    HDR+AA: Oblivion MountainsPage:: ( 20 / 25 )

    Oblivion – Direct3D






    Notes

    Under the greater demands of HDR+4xAA, the Radeon HD 2900 XT is able to pull a little closer to the GeForce 8800 GTS 640MB, but it still isn’t able to overtake it until 2560x1600. The factory overclocked board continues to remain ahead of the Radeon HD. This is really surprising considering AMD’s memory bandwidth advantage.

    We can clearly see that the alpha AMD driver does improve performance in Oblivion under HDR+AA though.



    HDR+AA: Oblivion FoliagePage:: ( 21 / 25 )






    Notes

    The alpha AMD driver continues to help the Radeon HD 2900 XT’s performance, in fact the card is able to outrun the e-GeForce 8800 GTS KO ACS3 at 2560x1600.



    HDR+AA: Far CryPage:: ( 22 / 25 )

    Far Cry – Direct3D








    8xMSAA PerformancePage:: ( 23 / 25 )







    Notes

    Even under 8xMSAA and 1920x1200 screen resolution, the Radeon HD 2900 XT isn’t able to overtake the GeForce 8800 GTS 640MB cards in FEAR or Lost Coast. This is the type of benchmark where you’d expect the Radeon’s memory bandwidth advantage to reap the most benefits, but it’s not. We did run some quick numbers with the Radeon HD 2900 XT and GeForce 8800 GTX in Company of Heroes with 8xAA and found the Radeon HD held an edge over the GeForce card at 1920x1200. Company of Heroes is certainly in app that the Radeon HD 2900 XT runs well in at the moment.




    Power, temps and OverclockingPage:: ( 24 / 25 )




















    ConclusionPage:: ( 25 / 25 )


    So why does such an impressive-looking GPU have such a tough time beating NVIDIA’s 6-month-old GeForce 8800 GTS in 3D performance? That’s a good question.

    Pulling up those paper specs once again, you can’t miss the lack of texture units. While shading performance is becoming increasingly important, shaders still depend on textures to do much of their work.

    Due to the nature of R600’s superscalar architecture, driver efficiency is also going to be more important. We have no doubt in our minds that the performance of the Radeon HD 2900 XT is being held up by the driver.

    Fortunately AMD’s driver team has a very solid history of delivering steady performance improvements as they become more familiar with the architecture. A month or two from now the performance picture could be drastically different than what we’ve presented to you today.

    But one aspect that no driver can fix is the Radeon HD 2900 XT’s power consumption. Now normally we don’t harp on aspects like power consumption, because honestly, as enthusiasts, we’d be willing to pay a few extra dollars per year on our utility bill if that means we’re getting cutting-edge performance. But clearly in the case of the Radeon HD 2900 XT, you’re not getting that. As of right now, the Radeon HD 2900 XT performs most similarly to a mildly overclocked GeForce 8800 GTS 640MB (if that) but consumes power like a GeForce 8800 Ultra.

    In addition, as it stands now the Radeon HD 2900 XT runs considerably louder than any GeForce 8800 card, and if ATI were to fix this problem by lowering the fan’s RPMs, the GPU’s power consumption would spike even higher. It’s a damned if you do, damned if you don’t situation that really hampers Radeon HD 2900 XT’s potential. There’s no way a Radeon HD 2900 XTX model can be produced until ATI solves their power consumption problem.

    Fortunately the Radeon HD 2900 XT has other things going for it that round out the package a little more. On the video side, AMD’s Avivo HD solution appears to be more flexible than PureVideo HD (video benchmarks from Alan coming shortly) while the ability to pass both audio and video via HDMI will really make setting up an HTPC a breeze. The addition of Half-Life 2 Black Box gives ATI a killer game bundle yet again. Another plus is overclocking – all early indications suggest that the Radeon HD 2900 XT has tons of frequency headroom for OC’ing. We were able to hit speeds 858MHz core/900MHz memory with both Radeon HD 2900 XT cards supplied by AMD, as well as the retail Sapphire Radeon HD 2900 XT we received. Anything beyond that and the slider in Powerstrip would revert back to 858/900 – we don’t know if we hit some sort of power ceiling (we conducted all our testing with a 1.1-kilowatt Tagan PSU so that seems unlikely) or if there’s some sort of artificial limit on how high you can OC the card.

    The only downside to overclocking is most users will have to upgrade to an 8-pin power supply in order to do it. Doh!

    Ultimately, the success or failure of the Radeon HD 2900 XT is going to be up to AMD’s driver team. Right now the card is hard to justify at $400, you can find GeForce 8800 GTS 640MB cards for less money at most online retailers. But if AMD’s driver team is able to squeeze more performance out of the Radeon HD 2900 XT, its $400 price tag becomes easier to live with. Especially once Half-Life 2: Episode Two and Team Fortress 2 are released…


    • © Copyright 2003 FS Media, Inc.
    [ Print Article! | Close Window ]