New games demand more of video cards and their memory; bigger textures, more triangles, and various other new means of hardware acceleration to look good and still run at playable speeds.
NVidia's chips aren't the only ones suffering from lack
of memory throughput, although theirs feel its sting most acutely. Several other PC
hardware sites have demonstrated that while overclocking the core of the GeForce2 GTS results
in modest speed improvements, overclocking the memory helps much more. Sadly, it's not easy
to overcome the limitation of memory speed. The problem is different from that
of main system RAM in several ways. First, the memory chips are integrated onto the video
card, and the physical distance between the chips and the video chip is rather short. (The
video chip itself contains the hardware which controls memory, so it
doesn't need a separate chipset like processors do.)
This means that they don't have to deal with bulky
memory modules, which cost more to fabricate and integrate because of the extra
PCB and connection mechanisms (e.g. DIMM slots). Each chipset can also have
basically any type of memory and datapath width that's practical, since
compatibility across multiple platforms isn't an issue. For example, the
GeForce2 GTS uses DDR memory on a 128-bit datapath, while 3dfx's Voodoo5 5500
uses two channels of 128-bit SDR memory, one per video chip. This allows the
memory subsystem of video cards to be less expensive to implement than main
memory, but it also needs much faster memory, so that offsets the reduced
fabrication costs quite a bit. It's still expensive to have really fast memory
with wide datapaths, but at least it's easier to do on video cards than with
main system RAM. Still, increasing bandwidth is not easy, and there are
basically four ways to do it: increase the frequency of the memory, use wider
datapaths, use multiple channels of memory (or multiple video chips, each with
its own memory), and use other techniques that increase throughput by other
means (like DDR).
Each of these methods has its drawbacks. High-frequency memory is very expensive and hard to find in quantity. Wider datapaths mean high pin counts and more complicated controller circuitry, which add to design complexity and fabrication costs, thus increasing the overall cost substantially. Using multiple channels seems to be easier than using wider datapaths, but it still adds considerably to the complexity of the design. New types of memory tend to be rare (and thus expensive) for a while, simply because they initially have small support bases and few, if any, chips or chipsets can support them.
I should note here that cost
is a big concern, more so these days than it used to be. While today's video
cards are much faster than those of last year, the newest ones are also much
more expensive than their year-old counterparts were when they debuted.
I remember paying $211 for my brand-new Diamond
Viper V770 Ultra, a TNT2 Ultra board with 32 megabytes of 183MHz SDR memory,
only last spring. Upon its release, the TNT2 Ultra was arguably the most
powerful desktop video chipset available. These days, the biggest contenders are
GeForce2 GTS cards with 32 and 64 megabytes of DDR memory and Voodoo5 5500 cards
with 64 megabytes of SDR memory. These cost upwards of $250, and the more
feature-laden GeForce2 GTS cards can be well over $300. Today's most powerful
cards, based on the GeForce2 GTS and carrying 64 megabytes memory, start at
about $330 for the cards themselves, while some full retail boxes can be over
$400. While $211 might not seem like too much to pay for one of the best video
cards around, Gamer Joe can't afford to dump $400+ on a new video card, and even
$300 is pretty steep for many. But increased performance demands faster and/or
more components, thus increasing cost and thus retail prices.
RSS Newsletter
