f 60 Hz, 150 KB of image memory had to be output to the screen 60 times per second, resulting in a memory bandwidth of just of over 9 MBps on the vide
o side.
The situation heats up considerably for the modern resolutions of SVGA and better. To accommodate a moderate 1024 by 768 pixels at 65,536 colors (16 bits per pixel), the image buffer has to hold about 1.6 MB of data. Thus, the CPU needs a bandwidth of 16 MBps to write data to the graphics board. This is no big deal on the PCI bus, which can handle up to 60 MBps. But to get the image out to the screen at an ergonomically correct 75 times per second, the board must read from its image memory at 120 MBps. For 1280 by 1024 resolution, you need a sustained throughput rate of 196 MBps.
How memory delivers this performance depends on the bus width used in the graphics card's design. Basically, more memory bandwidth can be achieved either with faster memory chips or by using several chips in parallel. ISA-bus graphics boards and their chips were generally designed for a data-bus width of 16 bits; to support high screen resolutions, you need fast video RAM (VRAM).
VRAM chips are du
al-ported: To the graphics chip, they look like standard DRAMs, but a second port is used for output. Thus, data can be read and written simultaneously; screen updates do not result in lost cycles. A recent development, Window RAM (WRAM), is basically faster VRAM enhanced with special graphics features, such as accelerated, aligned fills and moves.
Given that memory comes in a fixed bus width, watch out for boards that support a bus width of 64 or 128 bits but are sold in economy versions where only half the memory is fitted. For example, a 128-bit board might be designed for 4 MB of RAM, needing eight memory chips, each with a bus width of 16 bits. If only 2 MB is installed, which degrades the bus width to 64 bits, you effectively end up with a lower-performance, 64-bit board. Of course, you can install the missing memory later.
Modern boards, which transfer data at 64 or even 128 bits in parallel, can actually use slower, less expensive memory. For example, board designers can combine standard exten
ded data out (EDO) RAM with a 64-bit bus to achieve a throughput of 240 MBps. Burst EDO RAM delivers about 400 MBps.
Memory manufacturers have come up with a variety of special designs for even-faster throughput, needed for larger color palettes. Some designers put several chips in a single housing and combine them with logic circuitry that allows overlapping access. Each unit works at the speed of standard DRAM, but parallelism increases throughput. All the variations of DRAM described below are single-ported.
Synchronous DRAM (SDRAM) and synchronous graphics RAM (SGRAM) are available from several sources and have already found a flock of followers in the graphics-controller industry. While DRAM needs several special strobe signals, synchronous RAM can handle the memory clock directly. At a bus width of 64 bits, SDRAM can support bandwidths up to 640 MBps. SGRAM adds features that accelerate graphics performance, such as block write (which is useful for video).
Multibank DRAM (MDRAM) combines eig
ht 32-KB banks into blocks of 256 KB. It supports 500-MBps throughput. Tseng Labs' ET6000 graphics chip sports a special 128-bit MDRAM interface, yielding up to 1000 MBps.
Rambus DRAM (RDRAM), a general-purpose memory design, has a bus width of only 8 bits and therefore needs a very high memory clock of 250 MHz. Transfers reach 500 MBps.
Why Memory Matters
