Table of Contents
Course notes
Definition of Terms
Consumer PC graphics
Workstation PC
graphics
Exhibition
3D Graphics Chipset
Chart
3D Graphics
Accelerator Chart
3D Workstation
PC Systems Chart
Course notes
Course #29 - Developing High-Performance Graphics Applications for
the PC Platform
Definition of terms
Organizers of the course categorized PC graphic machines into two classes
termed "consumer" and "workstation". According to the organizers,
the dividing line between what is considered a "consumer" PC graphics machine
and a "workstation" graphics machine is the $300 mark for the graphics
accelerator card. Those PC machines with graphic accelerator cards
costing < $300 and in the general range of $3000 are termed "consumer"
PC graphic machines while those with cards > $300 are termed "workstation"
PC graphic machines.
Consumer PC graphics
The following material is from a talk by 3Dfx interactive (makers of
the VooDoo graphics board) founder and ex-SGI flight simulator developer,
Gary Tarolli.
Brief history
1994 - bleak for consumer-based 3D graphics
1995 - hardware texture mapping implemented
1996 - Voodoo card - 50M pixels/sec
1997 - nVidia Riva 128 - 1-5M tri/sec, full-scene anti-aliasing
1998 - Voodoo2 - 3M tri/sec, 90-180M pixels/sec, 2 simultaneous textures,
multi-texturing, texture compression
System configuration, performance
Games are now typically at 800x600 resolution or higher. Graphic
chips are "catching up" to CPU performance. Key has been semi-conductor
sizes (increased 40-fold from 50K gates on a graphics chip in 1992 to 2000K
gates in 1997).
Predictions
1999 - 5-10M tri/sec, bus bandwidth permitting, 300M+ pixels/sec, full-scene
anti-aliasing, texture compression < 4 bits.
2000 - 10M+ tri/sec, 900M+ pixels/sec
Innovation
In many areas, PC graphics have "copied" workstation ideas (i.e. z-buffer,
alpha, textures, fog, ...). However, PC graphic technology has "invented"
some interesting techniques and methods of implementing graphics:
- multiple textures
- low-cost full scene anti-aliasing
- high fill rates - low-cost
Conclusion
Not only is the line between workstations and "workstation" PC machines
"fuzzy" but now the line between "workstation" PC machines and "consumer"
PC machines has been dwindling. However, (quote) "PC cards will not
replace workstation PC cards."
What is it?
- Windows NT and multiprocessor support
- Optimized for OpenGL
- Sophisticated windowing support
- Higher pixel resolution
- Larger color depth
- alpha processor or intel
Brief history
before 1995 - 3DLabs GLINT 300sx, Dynamic Pictures (bought now by 3DLabs)
1997 - multiprocessor OpenGL driver, GLINT GMX, Oxygen RPM, 1.2M poly/sec
Exhibition
PC graphic machines contain 3D graphic accelerator boards, which, in
turn, contain 3D graphic chipsets. Before evaluating a PC graphics
system, one must examine the available 3D graphic accelerator board.
Before evaluating a graphics accelerator board, one must examine the underlying
3D graphics chipset. As a way of looking at systems, boards, and
chipsets, below are 3 separate charts listing 1) 3D graphic chipsets, 2)
3D graphic accelerator boards (using the chipsets), and 3) a few example
workstation PC systems (using the graphic accelerator boards).
As one could imagine, there is a myriad of 3D graphic accelerator boards
available. I tried to pick up all the brochures and gather information
from all the vendors represented at the exhibition, although I may have
missed a few. In addition, there are undoubtedly other boards and
chipsets not represented at the conference and hence do not appear in the
following charts. When drawing any comparisons, a minor warning --the
information (specs, performance, features) in the charts is directly from
the vendor brochures or web-sites and obviously contain performance test
specs and feature sets that mean different things to different
vendors. Finally, the information is no particular order.
Vendor | Product | Performance | API Support | Features |
3Dfx Interactive, Inc. | Voodoo2 | 3M tri/sec, 90M pixel/sec, Connect 2 Voodoo cards for double the fill rate and speed | 3Dfx Glide, OpenGL, Direct3D | Perspective correct texture mapping
Bi and tri-linear filtering Z-buffer LOD per-pixel MIP-mapping Sub-pixel and sub-texel correction Gouraud shading and texture modulation Full 24-bit rendering 14 texture formats, 8-bit compressed and 8-bit palletized formats 2 to 4 MB EDO DRAM frame buffers 2 to 4 MB EDO DRAM texture memory Anti-aliasing Depth buffering Alpha Blending Per-pixel special effects: fog, transparency, translucency Texture animation, morphing, compositing Full speed environment mapping Bump mapping Detail texturing |
3Dlabs | Permedia 2 | 1M polys/sec, 42M pixel/sec | OpenGL, Direct3D | Perspective correct texture mapping
Bi-linear filtering 16 bit Z-buffer Sub-pixel correction Gouraud shading Full-scene anti-aliasing Alpha Blending Stencil buffer Per-pixel special effects: fog |
3Dlabs | Permedia 3 | 8M poly/sec | Direct3D, OpenGL ICD | 32-bit color
32, 16-bit Z-buffer Per pixel tri-linear mip-mapping Full blend modes and alpha tests Anti-aliasing Bump mapping Virtual textures Voxel rendering Advanced filtering |
3Dlabs | Glint GMX | 2.6M visible, lit meshed tri/sec, 66M pixel/sec
GMX 1000: 1 Glint Gamma geometry processor + 1 MX rasterizer engine GMX 2000: 1 Glint Gamma geometry processor + 2 MX rasterizer engines |
OpenGL, Direct3D (DirectX geometry ready) | Gamma processor
100% OpenGL geometry pipeline in silicon Full geometry transform processing Full lighting calculations for 16 lights Texture coordinate and fog calculations Hardware clipping and culling Raster engine
24 or 32-bit Z buffer Multi-plane overlays, stencils
VRAM framebuffer
Screen resolutions up to 2048x2048 including HDTV resolution 100% hardware texture mapping (per-pixel mip-mapping, tri-linear filtering, on-chip texture cache, 8-bit palletized) |
3Dlabs | Oxygen RPM (previously of Dynamic Pictures) | 4M shaded tri/sec, 120M texture-mapped, mip-mapped pixels/sec
Power-threads for multi-threaded OpenGL |
OpenGL | Resolutions up to 1920x1200
24 bit true color and 24 bit Z-buffer Accelerated shading, lighting, texture-mapping Bi, tri-linear texture filtering Alpha blending with destination alpha support Multi-plane stenciling and overlay planes Anti-aliasing through multi-sampling Transparency, dithering, atmospheric effects |
Intergraph | Intergraph Intense 3D Wildcat | 6M shaded, lit, z-buffered, 25 pixel tri/sec, 90M pixel/sec
Provides hardware acceleration of all graphics functions via 3 custom ASIC devices (bus interface, geometry acceleration, rasterization). |
OpenGL, RenderGL, Direct3D, DirectDraw, 2D GDI | Texture acceleration (point sampled, bi, tri-linear mip-mapping, 32-bit
textures, 3D volumetric textures)
32-bit Z-buffer Per-pixel effects: complex fog, atmospheric effects Linear, exponential, user-defined fog models 8 stencil planes 8 overlay planes |
Evans & Sutherland REALimage 2000 | Evans & Sutherland REALimage 2000 | 4M primitives/sec, 90M pixel/sec | OpenGL, Direct3D | Full OpenGL feature set
Phong and Gouraud shading Alpha blending Anti-aliased lines and sorted polygons Atmospheric effects Depth cueing Bi, tri-linear mip-mapping Scissoring Subpixel processing Overlay and stencil planes Accumulation buffer Alpha-blended effects: fog
32-bit RGBA, double-buffer 24-bit Z-buffer |
nVidia RIVA TNT | nVidia RIVA TNT | 128-bit wide graphics engine and frame buffer
8M million triangles/sec peak, 250Mpixel/sec fill rate |
OpenGL, Direct3D | Optimized Direct3D acceleration
Complete DX5 and DX6 support Twin texel (TNT) 32-bit graphics pipeline Per pixel perspective correct texture mapping 24-bit or 16-bit HW Z buffer (floating point or integer) 8-bit stencil buffer Anti-aliasing: full scene, order independent High performance 128-bit 2D Acceleration Fast 32-bit VGA/SVGA support 16MB, 8MB and 4MB frame buffer configurations |
Intel740 | Intel740 | 1.1M tri/sec
45-55M pixels/sec |
OpenGL, DirectX | Flat and Gouraud shading
Bilinear filtering mip-mapping Color alpha blending for transparency Real-time texture paging and video texturing Fog and atmospheric effects Specular lighting Edge anti-aliasing Stippling transparency Backface culling Z-buffer Texturing (per-pixel correct, various formats, integrated hardware palette) |
Matrox MGA-G200 | Matrox MGA-G200 | OpenGL, DirectX | Alpha-blending
Bi, tri-linear mip-mapping Fog Anti-aliasing Specular highlights |
Vendor | Product | 3D Graphics Chipset | Memory | Price |
Intergraph | Intergraph 3D Wildcat 4100 | Intense 3D Wildcat | ||
Elsa | Gloria XXL | 3Dlabs Glint Gamma geometry chip | 16MB VRAM + 24/40MB DRAM | $2249 |
Gloria XL | 3Dlabs Glint MX + geometry chip | 16MB VRAM, 22MB DRAM | $1699 | |
Gloria L/MX | 3Dlabs Glint MX + geometry chip | 8MB VRAM, 16MB DRAM | $1099 | |
Gloria Synergy | 3Dlabs Permedia2 | 8MB RAM | $149 | |
Erazor II | nVidia RIVA TNT | 16/8MB RAM | TBD | |
Erazor | 128-bit engine | 4/8MB | $79 | |
Real3D StarFighter | Real3D StarFighter | Intel740 chipset | 4/8MB SGRAM | $83 - $220 |
AccelGraphics | AccelSTAR II | 3Dlabs Permedia2 | 8MB SGRAM | ?? |
AccelPRO MX | 3Dlabs Glint MX | 8MB framebuffer, 16MB local, texture | ?? | |
AccelGMX 2000 | 3Dlabs Glint GMX 2000 | 16MB framebuffer, 80 MB local, texture | ?? | |
AccelECLIPSE II | E&S REALimage 1000 | 16MB framebuffer, 4/16MB texture | ?? | |
AccelGALAXY | E&S REALimage 2000 | 20MB framebuffer, 16/32 MB texture | ?? | |
Real3D Oxygen RPM | Real3D Oxygen RPM | Oxygen chipset | 32, 64, 128MB | ?? |
Quantum3D | Obsidian2 Consumer | Voodoo2 chipset | 16/24 MB Memory | $320 - $700 |
Obsidian2 Professional | Voodoo2 chipset | ?? | ||
Diamond | Monster 3D II | Voodoo2 chipset | 8/12MB | $230 - $280 |
Fire GL 1000 Pro | 3Dlabs Permedia2 | 8MB SGRAM | $150 | |
Fire GL 3000 | 3Dlabs GLINT texture graphics+ GLINT geometry | 8MB VRAM to 32MB EDO DRAM | $800 | |
Fire GL 4000 | Mitsubishi + E&S Realimage | 15MB 3DRam | $2400 | |
Fire GL 5000 | E&S Realimage 2100 | 20MB 3DRam | ?? | |
Matrox | Millennimum G200 | Matrox MGA G200 | 8/16MB | $165 |
STB | Glyder GMX-1000 | 3Dlabs GLINT GMX1000 | 16MB VRAM, 40MB DRAM | ?? |
Glyder GMX-2000 | 3Dlabs GLINT GMX2000 | 16MB VRAM, 80MB DRAM | ?? | |
Leadtek | WinFast 3D L2520 | 3Dlabs GLINT MX | 24MB | ?? |
WinFast 3D L2530 | 3Dlabs GMX1000 | 24MB | ?? | |
WinFast 3D L3100 | 3Dlabs GMX2000 | 96MB | ?? | |
MaxVision | Titan II | 3Dlabs GMX2000 | 16MB VRAM + 80MB DRAM | ?? |
Some workstation
PC systems
For a flavor of high-end workstation PC's, below is a sample of a few
configurations with graphic accelerators, memory, and pricing. These
were chosen at random and obviously represent a much greater audience of
configurations and vendors.
Vendor | Processor | Graphic Accelerator | Memory | Price |
Hewlett-Packard Kayak | 400Mhz Pentium II, Dual-processor support | Visualize fx4 (HP) | $11,779 | |
NTSI | 400Mhz Pentium II, Dual-processor support | TitanII | 128MB - 1GB | $5000 - $10,000 |
Tri-Star Computer | 400Mhz Pentium II, Dual-processor support | ELSA Gloria XXL | 384MB | $11,500 |
400Mhz Pentium II | AccelECLIPSE II | 256MB | $6000 | |
400Mhz Pentium II | Oxygen | 128MB | $5000 | |
Core Microsystems | 350Mhz Pentium II, Dual-processor support | 3Dlabs GMX2000 | 128MB | $6695 |
Dell "Siggraph show special" | 400Mhz Pentium II | Intergraph Intense 3410GT | 128MB | $9000 |
Mitsubishi 3D-Ram
The new Mitsubishi 3D-RAM is an innovative frame buffer memory architecture
optimized for high-performance 3D graphics that provides approximately
3 to 4 times increase in rendering performance. Designed for high-performance
workstation and personal computer applications, the 3D-RAM increases the
data transfer rates by integrating key functions on-chip, including an
arithmetic logic unit (ALU), to accelerate Z-buffer rendering and on-chip
caches.
PixelFusion
In development of a series of processors using a termed they call FUZION.
FUZION combines massively parallel rendering and sub-pixel programmability
in a new scalable architecture. FUZION will be launched mid 1999
priced for high-end PC NT market. They claim the following:
11M tri/sec - peak
1 M anti-aliased, true phong shaded, bump mapped, mip-mapped tri/sec
- sustained
Anti-aliasing with sub-pixel Z-buffer
Programmable pixel shading delivering features such as true Phong shading
and bump mapping in real time
Optional geometry
OpenGL and Direct3D compliant
Advanced rendering techniques can be programmed (i.e. volumetric)
Real3D
PRO-1000
Image generators or graphics peripheral designed to be connected via
an SCSI II interface and driven by a user-selectable host computer.
The unit was formed as a separate entity to leverage and commercialize
real-time graphics technology developed by Lockheed-Martin for high-performance
combat simulators. Features, performance:
1M textured, shaded, translucent, anti-aliased, fogged poly/sec
built-in processing of fog and shading
illumination lobes
database culling
32MB texture memory