Upgrades in their architectures along with the integration of large buffer memories enable the new breed of graphics ICs to deliver supercomputer-like graphics performance on a single chip.
Today's computers can generate movie-quality 3D animation, a task that would have required days of rendering on multiple supercomputers only half a decade ago. To improve this far, many different technologies—high levels of chip integration, high-performance memories and host processors, and powerful graphics algorithms—had to join together.
The latest crop of high-end graphics engines packs close to 50 million transistors—even more than what some CPU chips have today, like Intel's Pentium 4 or AMD's Athlon families. The graphics chips are high-performance computers in their own right. They contain multiple pipelines to handle the pixel operations, as well as other computational units to perform shading, lighting, reflection, and other functions needed to produce amazingly realistic images and animation.
Many new laptop computers can deliver 3D graphics capabilities comparable to most desktops. That kind of performance is becoming a necessity rather than an option as many users view the laptop as a desktop replacement.
The once-crowded field of standalone graphics chips has condensed from over 30 suppliers to just a handful, with ATI and nVidia leading the pack. The survivors include Matrox, the S3 division of VIA Technologies, Silicon Integrated Systems, and Trident Microsystems. At the same time, manufacturers of PC motherboard chip sets are incorporating high-performance graphics engines right on the motherboard logic chip. The Intel 845G chip set is an example of that trend.
Nowadays, double-data-rate synchronous DRAMs (DDR SDRAMs) with clock speeds of 333 MHz and faster are used in frame buffers that deliver data bandwidths of 5 to 10 Gbytes/s to the graphics. The Rambus RDRAMs also provide a high-bandwidth memory interface but have met with limited success in the highly cost-competitive graphics market. Still faster memories are also being adopted—the 2G DDR SDRAM, with clock speeds starting at 400 MHz and increasing to at least 600 MHz over its lifetime, is starting to appear.
Today's host CPUs and chip sets support clock speeds of up to about 2.5 GHz, with even faster processors expected. Motherboard chip sets are also starting to offer AGP 8X interfaces to provide a 2-Gbyte/s interface between the PC and the graphics card. This will facilitate faster transfers of data into the frame buffer or other buffers on the graphics chip.
In addition to running the graphics-intensive applications, the CPUs handle many graphics support and interface functions. The combined compute power of the CPU and the graphics chip also leverages advances in software drivers, like Microsoft DirectX 9.0 and the latest OpenGL enhancements, to deliver more features and programming flexibility. To achieve supercomputer-like throughput, graphics-chip vendors have turned to extremely high levels of integration, combined with very wide databuses.
ATI Technologies' recent Radeon 9700 PRO graphics engine incorporates eight parallel rendering pipelines and four parallel geometry engines, a 256-bit wide DDR memory interface, and AGP 8X graphics support.
Those features coalesce with the company's Smartshader 2.0 technology, offering programmable pixel and vertex shaders with 16 textures per pass. The pixel shaders handle up to 160 instructions and can use floating-point precision levels of up to 128 bits for a greater range of colors and brightness levels. Vertex shaders can handle up to 1024 instructions with flow control.
ATI's Smoothvision 2.0 technology performs full-scene anti-aliasing and anisotropic filtering to improve image quality. Incorporated on-chip are a three-level hierarchical Z-buffer, as well as a video shader that seamlessly integrates the pixel shaders with video. The chip's Fullstream video deblocking technology delivers sharper looking images and packs noise-removal filtering to clean up captured video.
Archrival nVidia's recent family of graphics processors targets three market tiers: the GeForce 5800 for the high-end gamer, the GeForce 5600 for the mainstream user, and the GeForce 5200 for cost-conscious users. Each is based on the company's FX graphics-processor-unit architecture. The chips get their high-end performance from a rendering engine that works on eight pixels every clock cycle. The engine employs 128-bit floating-point precision throughout the graphics pipeline.
The FX architecture's Intellisample technology provides high-speed anti-aliasing, adaptive texture filtering, an advanced loss-less compression capability for both color and Z-data, and a fast Z-clear capability. An available Digital Vibrance Control helps sharpen images and offers enhanced color controls.
Based on this architecture, the FX 5600 delivers about 30% better performance at about half the price of nVidia's antecedent GeForce Ti4600. By removing some of the acceleration features, the FX 5200 graphics engine will let card vendors offer a full graphics adapter compatible with DirectX 9.0 for just $79.