The ATi Radeon 9000 and Radeon 9700 Preview
ATi 's Next Generation Products Unveiled

By, Dave Altavilla
July 18, 2002


 

Did we save the best for last?  This is certainly the star of the show for ATi, as it is their true next generation technology targeted at putting the hurt on NVIDIA's well entrenched GeForce4 Ti 4000 product line.  During a private meeting behind closed doors at one of New York's swank, marble trimmed hotels, we were treated to a hands on session with the Radeon 9700 that is powered by the ATi R300 VPU.  Certainly the paper specs that we'll show you here, were impressive.  However, what was more impressive was the actual in game performance we witnessed first hand.  Its fairly safe to say that ATi is going to deal out a "killer" type product with their R9700 board and it's coming as soon as next month.

First we'll cover the specifics of the R300 VPU (Visual Processing Unit) and then we'll tell you what we saw it do in at our meeting in New York...

Features of the Radeon 9700 Series
ATi's Fully DX9 Compliant Rendering Engine

GRAPHICS TECHNOLOGY

  • RADEON 9700 Visual Processing Unit (VPU)

MEMORY CONFIGURATION

  • 128MB of double data rate SDRAM

3D GRAPHICS FEATURES

  • Eight parallel rendering pipelines process up to 2.6 billion pixels per second

  • Four parallel geometry engines process up to 325 million transformed and lit polygons per second

  • High precision 10-bit per channel framebuffer support

  • 256-bit DDR memory interface

  • AGP 8X support

-SMARTSHADER 2.0

  • Full support for Microsoft DirectX 9.0 programmable pixel and vertex shaders in hardware

  • 2.0 Pixel Shaders support up to 16 textures per rendering pass

  • 2.0 Vertex Shaders support vertex programs up to 1024 instructions with flow control

  • New 128-bit per pixel floating point color formats

  • Multiple Render Target (MRT) support

  • Shadow volume rendering acceleration

  • Complete feature set also supported in OpenGL via extensions

-SMOOTHVISION 2.0

  • State-of-the-art full-scene anti-aliasing

  • New technology processes up to 15.6 billion anti-aliased samples per second for unprecedented performance

  • Supports 2x, 4x, and 6x modes with programmable sample patterns

  • Advanced anisotropic filtering

  • Supports up to 16 bilinear samples (in performance mode) or trilinear samples (in quality mode) per pixel

  • 2x/4x/6x full scene anti-aliasing modes

  • Adaptive algorithm with programmable sample patterns

  • 2x/4x/8x/16x anisotropic filtering modes

  • Adaptive algorithm with bilinear (performance) and trilinear (quality) options

  • Bandwidth-saving algorithm enables this feature with minimal performance cost

-HYPERZ III

  • Hierarchical Z-Buffer and Early Z Test reduce overdraw by detecting and discarding hidden pixels

  • Lossless Z-Buffer Compression and Fast Z-Buffer Clear reduce memory bandwidth consumption by over 50%

  • Fast Z-Buffer Clear

-TRUFORM 2.0

  • 2nd generation N-patch higher order surface support

  • Discrete and continuous tessellation levels per polygon for dynamic LOD

  • DirectX 9.0 displacement mapping

VIDEO FEATURES

-VIDEOSHADER

  • Seamless integration of programmable pixel shaders with video data

  • High quality, hardware accelerated de-blocking of internet streaming video

  • Noise removal filter for captured video

  • Integrated MPEG-2 decode

  • Hardware accelerated iDCT, motion compensation, and color space conversion

  • Top quality DVD and all-format DTV/HDTV decode with low CPU overhead

  • Back-end scaler delivers top quality playback

  • Upscaling and downscaling with 4-tap horizontal and vertical filtering

  • Filtered display of images up to 1920 pixels wide

  • Unique per-pixel adaptive de-interlacing feature combines the best elements of the bob and add-field (weave) techniques

FULLSTREAM video de-blocking technology

  • Noise removal filtering for captured video

  • MPEG-2 decoding with motion compensation, iDCT and color space conversion

  • All-format DTV/HDTV decoding

  • YPrPb component output

  • Adaptive de-interlacing and frame rate conversion

  • Dual integrated display controllers

  • Dual integrated 10-bit per channel 400MHz DACs

  • Integrated 165 MHz TMDS transmitter (DVI and HDCP compliant)

  • Integrated TV Output support up to 1024x768 resolution

  • Optimized for Pentium 4 SSE2 and AMD Athlon 3Dnow!

  • PC 2002 compliant

DISPLAY FEATURES

  • Dual integrated display controllers

  • Drive two displays simultaneously with independent resolutions and refresh rates

  • HYDRAVISION software provides complete control over multi-display configurations with a user-friendly interface

  • Dual integrated 10-bit per channel palette DACs operating at up to 400MHz

  • Integrated 165MHz TMDS transmitter supports resolutions up to QXGA (2048x1536) and complies with DVI and HDCP specifications

  • Integrated TV-Out support up to 1024x768 resolution

  • YPrPb output for direct drive of HDTV monitors

DISPLAY SUPPORT

  • 15-pin VGA connector for analog CRT

  • S-video or composite connector for TV/VCR

  • DVI-I connector for digital CRT or flat panel

  • Independent resolutions and refresh rates for any two connected displays

GENERAL FEATURES

  • Comprehensive 2x, 4x, and 8x AGP support

  • High performance quad-channel DDR memory interface supports 64/128/256MB configurations

  • Fully compliant with PC 2002 requirements

  • Optimized for Pentium 4 SSE2 and AMD Athlon 3Dnow! processor instructions

  • Supports optional THEATER 200 companion chip for NTSC/PAL/SECAM video capture

  • Highly optimized 128-bit 2D engine with support for new Windows XP GDI extensions

Although ATi didn't have any review samples of the 9700 available, we were able to get a snapshot of the board as well as a list of some of its features.  We were told that boards should be available some time in August.


Let's go through some of  the salient points and features of the all new Radeon 9700 and the VPU (now referred to by ATi as a Visual Processing Unit) that powers it, the R300.

The Hardware:
Physically, the Radeon 9700 is an impressive display of semiconductor technology and board design.  The GPU itself is targeted at running at a core clock speed of 325MHz.  This is impressive in and of itself, especially when you think about the Matrox Parhelia.  The Parhelia has "only" 80 million transistors, is also built on a .15 micron process, but has a clock speed of only 220MHz.  The R300 has 107 million transistors, so the chip is actually roughly 25% larger than the Parhelia.  The Radeon 9700 card is not much larger than a R8500 board, perhaps a bit longer but well within the AGP spec.  It does however have a 4 pin power connector on the back end of it, just like a floppy drive power plug.  This is for certain system configurations that may have problems supplying enough power through the AGP interface. 

ATi wouldn't comment at the time, as to how many PCB layers the R9700 board has.  However, with the 1000+ ball BGA package of the R300 VPU, routing all those connections is certainly not trivial.  We estimated it to be a 6 or 8 layer board.

The R300 VPU Engine:
Core Clock Speed Target - 325MHz
Memory Clock Speed Target - 310MHz (620MHz DDR)
.15 Micron Technology - 107M Transistors
Full AGP 8X Support

Obviously, if you scan through the above features list, this is a totally new architecture with dramatically more raw horsepower than ATi's legacy R200 chip.  The first thing that should jump out at you is the fact that this VPU has 8 parallel rendering pipelines and 4 parallel geometry engines. This bring forth a blistering 2600MPixel (or 2.6GPixel) Fill Rate.  ATi also went for high precision rendering with the R300 as it supports 10-bit per channel color in it's frame buffer.

SmartShader 2.0:
This is the heart of soul of the R300.  It has fully DX9 compliant Pixel and Vertex Shader Engines that can process up to 16 textures per pass, versus the R9000 and R8500 at 6 textures.  The real kicker here is now these shaders can now process color formats with floating point precision.  This means full 128 bit color precision on a per pixel basis.  DX9 requires 64 bit precision specifically, so this is more than enough for compliance and additional shader detail that developers may be able to utilize.  There are a host of other features in ATi's SmartShader 2.0 technology, like Shadow Volume rendering acceleration but the real special sauce is it's floating point precision and the number of textures per pass that it can handle.  This new engine should be able to perform orders of magnitude faster than the Radeon 8500 or GeForce4 Ti 4600, especially when handling next generation game engines like Doom 3 or Unreal 2003.

SmoothVision 2.0:
ATi has also revamped their AA and Anisotropic Filtering technology as well.  This AA technology is a "Multi-Sampling" approach versus the Super-Sampling method of the Radeon 8500 and 9000.  We don't know too much more about this technology, except to note that it is definitely a Full Scene AA approach and the GPU is able to process this, once again at an order of magnitude faster than previous generations and at higher resolutions.  You'll also note that in addition to 2 and 4X AA, there is now a 6X AA mode available as with the R9000.  One final note on AA is that the R300 VPU also supports sample gamma adjustments, which will actually provide a smoother looking AA at the edge pixels than traditional AA methods.  We saw 6X mode running and it looks absolutely gorgeous. We'll comment on frame rate here later in the piece but it is safe to say, we were impressed with that too. 

Finally, Anisotropic Filtering capabilities have also been enhanced and ATi's "selective filtering" techniques, which only filter the textures that need it most in a given scene.  The 9700 now supports up to 128 tap or 16 trilinear samples, in there "performance mode" driver settings.

HyperZ-III, TruForm 2.0, Video Engine, DACs,
and Performance Commentary