Have a look at the preview of a prototype Gainward CARDEXpert GeForce2 GTS/400 on Tom's Hardware Guide and notice the huge increases in speed at high resolutions with the memory overclocked. They overclocked the core to 233MHz, but the clear cause of most of the speed boost is the memory. In other tests I've seen where the core alone was overclocked, speed gains tended to be in the low single-digit percentages even if the card was clocked up to 240MHz. The card uses 5ns DDR memory (rated for 200MHz) but, as you can see, Tom managed to crank it up to 236MHz (effectively 472 MHz), and the speed gains are impressive at that memory frequency. If you don't want to read through the article, here are some of the numbers I found most impressive. Running Quake 3 Arena Demo 001 at 1600x1200 with Normal quality settings at 16-bit color, a 64-megabyte GeForce2 GTS with 166MHz DDR memory turned in a score of 64 frames per second. The overclocked Gainward card (233MHz core, 236MHz DDR memory) scored 87.7 FPS. That's a 37% increase. Running the same demo at High Quality 1024x768 (32-bit color), the standard card got 49.3 FPS while the Gainward got 70.6. That's a 43% increase from a 42% increase in practical memory throughput and a 16.5% increase in core speed.

Clearly, this card is mostly limited by its memory speed at very high resolutions.
The GeForce series of chipsets, as well as Matrox's and many of ATI's cards, all use just one core video chip, as most companies have for years. As technology and designs improve, they just cram more and more features onto those individual chips. But 3dfx's approach is different. Their newest cards use multiple VSA-100 video chips that work in tandem, and presumably their future generations of cards will use multiple chips as well. ATI has also dabbled in multiple-chip cards with its Rage Fury MAXX, which uses dual Rage 128 Pro chipsets, and it reportedly has plans for another dual-chip card that will use two Radeon 256 chips.

As you might expect, the primary benefit of using multiple chips is scalability. If each chip has the necessary hardware on it to cooperate with other chips of the same type, cards with extremely high fill rates can be made. There are quite a few drawbacks to this method, though. In such multiple-chip cards, each video chip needs its own bank of memory, since due to the way the chips must cooperate, a lot of data has to be duplicated in each bank. Thus, the 64 megabytes on the Voodoo5 5500, which uses two VSA-100 chips, aren't as effective in terms of storage space as 64 megabytes on a single-chip video card, because so much of their space is used for copies of data. But another advantage of this method is that the effective bandwidth is doubled. Because you have two memory paths working in tandem on a Voodoo5 5500, the total bandwidth is practically the same as dual-channeled memory. Of course, component cost is an obvious and significant drawback of this method.

The Voodoo5 5500 uses two core chips and two banks of memory, and that doesn't come cheap. In addition, the coordination logic on each chip adds a little more to its cost, as does all the other hardware that deals with controlling multiple chips. The video card also needs to be physically bigger to have room for all that stuff, and that increases the price further. Lastly, those two chips need a lot of electricity- so much that 3dfx thought it prudent to add a power connector (for use with a drive power plug) on the V5 5500 to make sure that the card gets enough juice. That's not to say that power isn't an issue for single-chip cards; GeForce 256 DDR cards had power requirements very close to the maximum that the standard AGP port can supply. But the problem gets worse as more chips are added. 3dfx plans to release a titan of a video card, the Voodoo5 6000, sometime this summer.

It will sport 4 VSA-100 chips and a whopping 128 megabytes of memory (again 32 megabytes per chip). The pictures of it I've seen remind me of the old full-length MCA and ISA video cards from days of old. And, not surprisingly, it will be a huge electricity guzzler. It needs so much power, in fact, that it will come with its own separate power supply. Personally, I think that's a very bad sign. Video cards shouldn't need their own power supplies. It's also expected to cost some $600, so it will be out of the price range of almost everyone. Quantum3D, a company which uses 3dfx's chips for huge (and extremely expensive) multi-chip cards, is planning to make a card with 32 VSA-100 chips and 64 megabytes of memory per chip (2 gigabytes total).

I'm forced to wonder how exactly they'll connect it to the motherboard, since seems unlikely that they could fit that many memory and VSA-100 chips onto even a full-length AGP card. Plus, it'll need 1600 watts of power, so it'll definitely need its own power supply. Anyway, it's not going to be used by anyone except companies dealing with high-end modeling, but it serves as an example of how scalable multi-chip technology can be.

While ATI plans to produce more dual-chip cards, they will be secondary to their lines of single-chip cards. It seems unlikely that any companies besides 3dfx will use multi-chip technology extensively for a while. The other companies will probably concentrate on adding new features and using smaller fabrication processes. Using a smaller fabrication process is one of the best methods of increasing the effectiveness of a component. Not only does it reduce the size of the component (which decreases cost), it also allows the component to run at higher frequencies (increasing speed, of course) and use lower voltages (thus reducing power requirements). 3dfx chose to use an already reliable enhanced .25-micron process with its VSA-100 chips, which is, depending on how you look at it, most of a generation behind everyone else, as they're using .18-micron processes. My guess is that they didn't want to risk moving to an untried .18-micron process, since the VSA-100 cards were already so far behind schedule. Unfortunately, this resulted in bigger, power-hungrier chips, which means higher overall costs and thus lower profits for 3dfx.

< Previous Page

© 2023 PCSTATS.com

Next Page >