This article is from the comp.sys.ibm.pc.hardware.video Frequently Asked Questions, by Michael Scott with numerous contributions by others. (v1.0).
This is one of the most commonly asked question in this group, and is usually answered more or less correctly, though often for the wrong reasons.
DRAM (Dynamic RAM) used on video cards is the same technology as the main system RAM on most computers. The 'dynamic' part refers to the fact that this type of memory must be refreshed several times per minute or it will 'forget' the data it is storing. This means that DRAM has a duty cycle (a period during which the RAM is being refreshed and can't respond to external requests like reads/writes), unlike SRAM (Static RAM) which does not require refreshing, and thus is available at all times. DRAM, however, requires fewer discrete components for each bit stored, so physically takes less silicon, and thus is cheaper to manufacture.
An additional limitation of DRAM is that it can do only one thing at a time - it can either be read from or written to. There are two data transfer steps occurring on your video card. The first is to transfer data from the CPU to video RAM. The second is to transfer the video RAM data to the RAMDAC, which produces the video signal you see on your screen. The maximum amount of data which you can pump in and out of your video memory in one second is your 'video bandwidth'. Thus, the read and write operations must share the available video bandwidth, which means that the DRAM has to service both read requests from the RAMDAC and write requests from the CPU. At high pixel addressabilities and colour depths, an enormous amount of extra data has to be moved to and from the video memory, and as a result, DRAM boards may run out of bandwidth. This means that you may not be able to refresh your monitor fast enough to avoid flicker.
VRAM is a special type of DRAM which is dual-ported. It still has a duty cycle, but it can written to and read from at the same time. In practice, this means that you get double the bandwidth out of 60 ns VRAM as you would out of 60 ns DRAM (if implemented correctly on the video card).
The long and the short of this is that VRAM cards are capable of higher screen refresh rates at high pixel addressabilities and colour depths, while DRAM cards are not. Some VRAM cards provide marginally better performance than comparable DRAM versions at lower addressabilities, but this will not affect the majority of users significantly. It will affect you if you run your monitor at high pixel addressabilities _and_ colour depths. Typically VRAM based cards perform better where DRAM cards drop off; noticeably at pixel addressabilities and colour depths greater than or equal to: 800x600 x 24 bit colour (16.7 M colours) 1024x768 x 16 bit colour (64k colours)
DRAM cards may be unable to provide high vertical refresh rates (>70Hz) at higher addressabilities. Most people aren't bothered by refresh rates >=60 Hz.
For techies who are looking for more detail: [From: Sam Goldwasser (sam@stdavids.picker.com)]
Both types are used to store video information. However, VRAM is not just fast DRAM. In fact, the random access times for typical VRAM is worse than for similar DRAM:
VRAM is a special type of DRAM which includes a shift register that can be loaded in parallel from an entire row in the DRAM array in approximately the same time as a single read cycle. The shift register (typically 256-2048 stages depending on the organization of the memory array) can be clocked independently of the normal random access to the chip.
The original intended use was for refreshing raster scan displays - thus the 'V' for video. Since the shift register is clocked independently, the percentage of time that the VRAM random access port is busy servicing video refresh is reduced from 'very high' to almost insignificant. For example, using DRAM, a typical design may require 50% of the random access port bandwidth for video refresh with DRAM but only .5% with VRAM. You load the shift register only once or twice per video line rather than having to access the memory array for every pixel. Some designs have a split shift register which provides even more flexibility in shift register load timing.
VRAM is slightly more expensive on a $/MB basis and is usually about 1 generation behind in terms of common chip densities. 4 Mbit VRAM chips are just now becoming commonplace.
There are a number of variations on this basic theme including some triple port varieties as well.
In addition to video, VRAM finds application in high performance printers and plotters, communications, signal processing, image capture using the shift register for input), and many other areas.
Also, from: hhanemaa@cs.ruu.nl (Harm Hanemaaijer):
Each resolution takes up a certain amount of bandwidth for monitor refresh. If this takes up most of the available bandwidth, performance goes down steeply.
With VRAM the bandwidth for drawing is basically unaffected by monitor refresh.
Resolutions where this happens are
bandwidth left 1024x768x256 NI on a 1Mb DRAM card 45 Mb/s 800x600x16bit on a 1Mb DRAM card 20 Mb/s 1Mb VRAM card (all resolutions) 100 Mb/s
(1Mb DRAM card has 60 MHz MCLK yielding 120Mb/s of memory bandwidth)
1024x768x256 on a 64-bit 2Mb DRAM card 165 Mb/s (good) 1024x768x32K on a 64-bit 2Mb DRAM card 90 Mb/s 800x600x32bit on a 64-bit 2Mb DRAM card 40 Mb/s 64-bit 2Mb VRAM card (all resolutions) 200 Mb/s
(2Mb DRAM card has 60 MHz MCLK yielding 240Mb/s of memory bandwidth)
It follows that so called 64-bit DRAM cards with only 1Mb are a pretty bad idea.
It can also be seen that 2Mb 64-bit DRAM cards can be faster than VRAM in very low resolutions that take up little bandwidth since the total bandwidth of the DRAM card may be a bit higher (e.g. 240 vs. 200Mb/s).
I'm not sure about the typical bandwidth of VRAM-based cards, but as far as DRAM cards are concerned most aggressively timed S3-864 based cards it is 120 Mb/s (1Mb) or 240 Mb/s (2Mb) while for more conservatively timed cards (which may imply better stability) it is about 100 / 200 Mb/s (this also goes for most CL-GD5434 based cards). You might imagine the performance vs. stability dilemma faced by manufacturers on this issue (the conservative 1Mb model has only 25Mb/s bandwidth at 1024x768x256 -> bonehead tester thinks it sucks).
 
Continue to: