Generally, an on-board accelerator chip can transfer graphics data between itself and banks of dual-ported VRAM (video RAM) faster than it can with DRAM, and faster over a 64-bit data path than a 32-bit path. The downside is that VRAM is more expensive (at least 1.5 to 2 times DRAM costs). As a compromise, graphics accelerator chip manufacturers such as Tseng Labs and Weitek use interleaving techniques on 32-bit-wide data paths to increase data transfer speeds between their accelerator chips and a graphics card's memory.
In short, memory interleaving is similar to double buffering. That is, while one bank of video memory is sending graphical data, a second bank can be receiving incoming data. Although you can also apply memory interleaving on VRAM-equipped cards, it is more typically used with DRAM because both fit with cost-conscious design goals.
On the other hand, manuf
acturers like ATI, Cirrus, Matrox, Media Vision, S3, and Weitek (again) have introduced accelerator chips that use a 64-bit-wide data path to move data between the graphics chip and video memory, and even sometimes out to the color DAC (D/A converter) that pumps the analog signal to the monitor. At this writing, controversy is brewing over whether a 32-bit interleaved design is as fast as a straight 64-bit-wide on-board data-path architecture.
Companies such as Tseng Labs that produce accelerator chips with 32-bit interleaved designs contend that a properly designed 32-bit interleaved accelerator card is every bit as fast as a 64-bit-wide accelerator card. My tests with three Tseng-based cards (from ATI, Diamond, and Genoa) show that they just can't keep up with 64-bit cards. The results for the fastest of the three cards, Genoa's Phantom 32i, are included along with the data from the 64-bit cards in the table "Performance for 64-bit Graphics Accelerator Cards" on page 138.
A Bigger Bus
