Each memory chip is a matrix of tiny cells. Each cell can contain one bit of information. The memory chips are usually called by the amount of information can keep qu. We call this chip density. May have found an example of the type of chip densities "64Mbit SDRAM" or "8M by 8". A 64Mbit chip contains 64 million cells and can hold 64 million bits of information. The expression "8M by 8" describes a kind of 64Mbit chip in more detail.
In memory industry, DRAM chip densities are often described by their cellular organization. The first number in the expression describes the depth of the chip (in locations) and the second number indicates the size of the chip (in bits). If we multiply the depth by the size, we obtain the density of the chip.
The off board 3D cards also include a generous amount of video memory (512 MB or more on newer models), accessed through a fast bus. The GPU (the chipset) is also very powerful, so that the two things combine to offer a performance monster.
With the introduction of PCI Express, also came the ability to install two or even four cards connected in SLI or Cross Fire (in the case of plates AMD / ATI), which provides a performance close to double (or quadruple) obtained by a single plate.
Away from the bright world of high performance boards, the boards have onboard, which are by far the most common. They are much simpler solutions, where the GPU is integrated into the chipset itself and the motherboard, rather than using dedicated memory, as the plates off board, uses part of main RAM, which is "stolen" from the system.
Even an old card like GeForce 4 Ti4600, has 10.4 GB / s bus to video memory, while using a stick of PC 3200 DDR memory, we have only 3.2 GB / s bus to main memory, that for a up needs to be shared between the video and the main processor. The processor handle it well, thanks to the L1 and L2 caches, but the video card actually has no where to run. That's why the onboard video chipsets are usually much simpler: even an expensive and complex chip would offer much better performance as the greatest constraint is the memory access.