Why nVidia is better
May 04 '00 (Updated Jul 22 '00)
3dfx staff has a major problem. Anytime you read an interview with them, they're always whining that graphics technology being implemented now isn't useful yet, and therefore their own products, which are designed around present game technology, are better. Ok, so you're the type of person that spent $600 on a couple of Voodoo2's back in the day so you could have SLI. This opinion is not for you because you would buy a new card every 6 months and are clearly insane.
With all that said, 3dfx ultimately loses this battle because they have not future proofed their product. With the GeForce and GeForce2, nVidia is turning hardware transform and lighting into a mainstream software feature. Add in a sort of functioning T&L engine on the s2000, and the ATi Radeon coming out next month, and you can be rest assured that a year from now hardware transform and lighting WILL be an important feature in the best games, which of course will be the ones us 3d speed junkies will want to play. Even 3dfx themselves have admitted that they will incorporate hardware transform and lighting into their chips "when the time comes". So if 3dfx is telling you that this technology will be important in the future, why would you let them sell you a product without it?
The reason I say a year is, hey, we're now looking at a hefty $300-400 investment for the best video cards out there. We deserve a good return on investment, and that doesnt include needing to buy a new product in 6 months for another $300. These products SHOULD last us a couple years at least.
So what does nVidia have thats so special? Not only does it have T&L that now handles up to 25 million triangles/sec, but it also happens to have the fastest chip on the market at the moment, which is resulting in the fastest video in the retail market. You can argue "well, when the voodoo5 6000 comes out...", but it's not out, and there's even a chance that it never will be, and hang onto your wallet if it ever is. So we have 800 megapixels/sec fillrate combined with a good T&L engine. Look at the benchmarks for the special nVidia Quake3 map(if you havent seen it or don't know what I'm talking about, check out this link: http://www.anandtech.com/reviews/video/nvidia/geforce2gts/nv15level.jpg Holy Schnikes, Batman!) with the super-high polygon count. It looks great and topped 30 fps on GeForce2's and Radeon's at decent resolutions. I wish games looked like that now, and I'm sure they will in a couple years. But just try running that on a Voodoo5 and you'll see how disappointed you are.
And therein lies my point: 3dfx is selling products that might get you 60fps in Quake3 now, but what will it do for me lately next year? Try raising that polygon count, load up the cpu with physics, and download the T&L onto the video processor, and the Voodoo5's weakness will come to light because it can't do hardware T&L. The sad truth is when that day comes, the Voodoo5 will fail miserably, and I, for one, am holding onto my video card for a while if I'm spending $400 for it. All I can say is pray for Rampage, and hope the company survives the vsa catastrophe.
A look into the future: 3dfx is clearly behind the game at this point, not too many people will argue that, even die-hard Voodoo fans have to admit that the GeForce 2 and ATi Radeon are both faster. Gigapixel poses some interesting possibilities with tile based rendering, but they're not looking at truly implementing any of that technology for another 2-3 generations, and by then, who knows what video hardware tech is going to look like. To make matters worse, STMicro will soon be releasing a new tile-based rendering card, the Kyro, which should offer about the same performance as a Voodoo5 for half the cost. Rampage better include a T&L engine, or 3dfx will dig it's own grave. From the rumors I've read, Rampage looks impressive, and if nothing else, promises to give nVidia a lot better run for their money than the Voodoo5 is.
I pretty much called it when the GeForce came out. When people were complaining about the low core speed, and the texel rate being the same as the pixel rate, I realized what nVidia was doing. They were optimizing their return on the research investment by designing a chip that was superupgradable. I was saying back then that they would make a TnT version of the GeForce, and sure enough, we're staring 1.6 gigatexels in the face. DDR RAM was just coming out in bulk, I put 2 and 2 together, and sure enough... The only thing I didnt see coming was the NSR.
Here's what I see in nVidia's immediate future: The Geforce2 will continue to reign supreme among hardcore gaming cards, at least in 16-bit. nVidia has a memory bandwidth problem which is evidenced by the severe performance hit when switching to 32-bit rendering. This is their number one problem to solve. This is made blatantly obvious by some benchmarks Tom's Hardware did of a prototype GeForce2 board that had it's memory overclocked to 236mhz as opposed to the reference board's 166mhz. Under those circumstances, that Geforce2 would easily surpass the soon-to-be-released Radeon, even in 32-bit. The only problem is, there is a big squeeze on high speed memory right now, which is serving as a bottleneck for just about the entire industry. There are a couple of ways around the memory bottleneck. One is tile-based rendering which only renders a section of the screen at a time, and later combines them. The other is to mathematically remove objects which can't be "seen" onscreen so that the video card has less to render. ATi is implementing this method through what they call HyperZ. nVidia supposedly will be incorporating similar technology in their next cards, coming out around september, but they call it "hidden surface removal". From a raw speed point of view, nVidia is king for the moment, and probably a while, but their dramatic advantage in horsepower won't mean much if they're straddled with the same memory bandwidth as everybody else. Most video card companies claim to be working on solutions to this problem, but they all refuse to give specifics, so I guess we'll just have to wait and see how it plays out.
With the ATi Radeon coming out, and the Voodoo5 appearing to be a disappointment, nVidia must now turn their attention to staying diligent in their performance increases. It is imperative that the nv20, coming out in a couple of months, be significantly faster than both the Geforce2 and Radeon. After that nVidia must ensure that their spring refresh surpasses 3dfx's Rampage, which is supposedly targeted for release around christmas. Also, it will be a good idea for nVidia to keep an eye on ATi, because they are now offering a very good product based on one chip. Not only will their refresh show up just a couple of month's after nVidia's, but they have an ace up their sleeve in the possibility of a dual Radeon card, which could be difficult to beat.
nVidia's TNT2 Ultra was capable of about 8 million triangles/sec, whereas the GeForce2's full potential is probably about 30 million triangles/sec. Thats a one year increase rate of 375%. We can use this rate to approximate the T&L performance on some of nVidia's upcoming products. The initial fall release nv20 should be 375% faster than the Geforce's 15 million triangles/sec, or around 56 million. Using the same increases, we can estimate that the spring refresh of the nv20 might be around 110 million triangles/sec. At 60 million tps, the nv20 would clearly kill the competition, but would be behind the planned 75 million tps of the Rampage at christmas, but would once again overcome that in the spring, just like it needs to.
Keep in mind that all PC graphics accelerator manufacturers now have to step up to the plate and compete with the PlayStation 2's 75 million triangles/sec. It won't suprise me if the PlayStation 2 is well ahead of the PC market when it's released. Then look for PCs to regain the lead again at christmas or next spring. The same scenario is probably not going to be true at the time of the planned release of the X-Box and it's theoretical projected acceleration of 300 million triangles per/sec, which, as we know, will contain an nVidia chip. According to prevailing performance increases, those kinds of figures are possible, but nVidia would have to follow a VERY tight schedule to achieve those numbers. And that, which anyone who has followed 3d the past few years knows, doesnt usually happen.
I once read an article that said photorealistic 3d acceleration would require about 90 million triangles per scene, or approximately 3-4 billion triangles per second. Realistically, current video processors can handle about 150,000 in a given scene.
If you do the math with current trends in processing power, that could be as little as 5 years away. Now if we could only solve that memory bandwidth problem...
 |
|
|
|
|
|
|
Epinions.com ID:
clburdette
|
|
Member: Curtis Burdette
Location: california
Reviews written: 46
Trusted by: 19 members
About Me:
question everything
|
|
|
© 1999-2009 Shopping.com, Inc.