|3d Update: Hardware T&L
by Leonard "Viking1" Hjalmarson
Last week S3 announced its next-generation graphics platform, the Savage2000. Shortly after Nvidia released information on their GeForce 256 I posted a 3d update that was short on information but long on speculation! Hey, what's a guy to do?
After many emails (special thanks to Paul Cross at S3 and to John Reynolds, hardware wizard and friend), and many forum responses, I want to correct some errors and share a little more clarity with our readers.
My first and largest mistake, noticed only by a select few (!) was to attempt to distinguish between an onboard geometry processor and the T&L engine. There is no distinction: they are one and the same. The new Nvidia product and the Savage2000 both sport an integrated geometry processor. The distinction between the two products is not clear, in terms of raw horsepower, except that the Savage2000 fill rate is greater than the GF256.
The geometry processor is the hardware that accelerates the T&L work by offloading it from your system's CPU. And this is separate from the fill rate issue, which is a claimed 700+ Mtexels/second for the Savage2000, twice the fill rate of the current boards. Some have estimated that the geometry processor on the GeForce 256 is 10 to 15% more powerful than that of the S3 part.
Fill Rate vs. Polygon Power
Consider for a moment that NVIDIA has more transistors in this chip than the next generation "Merced" CPU from Intel (23 million to be exact) and that they implemented what they call a "4x4 architecture." All this means is that each of the four key steps in the graphics pipeline are no longer done by the CPU, but rather by the GPU on the GeForce 256. Transform, lighting, triangle setup and the rendering will be done onboard with both the Savage2000 and the GeForce.
Polygons per second increases dramatically with hardware transformation and lighting - around 4 to 5 times what you would get with your CPU doing this task. The result is scenes that are much richer in detail. The GeForce is rated at 15 million polygons/second; S3 has not yet released a figure for the Savage2000.
What is unclear is how the moderate improvement in fill rate relates to the raw polygon power of onboard geometry processing. But here is the scenario as I understand it.
Where a game supports either OpenGL or DX7 (hardware T&L has been supported in OpenGL since 1.2), the performance increase will be astonishing. By offloading the CPU significantly, we will see increased detail at the same time as we see increased frame rates.
However, with older DX5 and DX6 games, only the increased fill rate will apply. So, with European Air War, for example, the GeForce at 120MHz clock might only give the gamer a 25-40% increase in frame rate over a TNT2.
The low-end fill rate noted by Nvidia for their GeForce 256 is 480 M/texels per second, barely 30% faster than the current reigning champion: the Voodoo 3 3500. The pixel fill rate is 2.5 times that of the 3500! Let's guess this truly is low ball, since the low end clock rate is a mere 120MHz. Guillemot has announced that their 3d Prophet, based on the same chip, will exceed this figure. Let's guess that Guillemot's board will top out at 150MHz. An interesting fact about the GeForce is that fill rate increases twice as fast clock for clock compared to TNT2. Moving to 150MHz would provide a fill rate of 600 Mpixels/Mtexels per second.
S3s Savage2000 similarly boasts a fill rate of 700 M/pixels per second, almost 400% higher than the V3 3500. This raw difference in horsepower means that most 3d games would see a dramatic increase in speed apart from other factors, including improvements in CPU clock rate or onboard T&L.
P51 in B17 II
On the other hand, the enormous T&L power GeForce chip means that as 3d games increase in polygon count and lighting complexity, you will notice increased benefits from the GeForce 256. No longer will users be turning off light sources for increased frame rate; additional sources simply won't impact the frame rates of most games running under the Savage2000 or the GeForce 256 in DX7.
The 3dfx Alternative
Over the last few months I've received anonymous mail detailing 3dfx new hardware a couple of times. Both times the mailer (let's call him "Mail-Throat") has given me information which later proved accurate. After my article on the GeForce went up last week I received more mail from my source claiming that the 3dfx product we would see this fall would be at least 2.5x the fill rate of the V3 3000.
This opinion is corroborated, if not understated, by two other sources. First, Bubba Wolford's report on the 3dfx technology demo last month; and second, by an interview with 3dfx that appeared on FiringSquad on the weekend. 3dfx strategy clearly focuses on the T-buffer (and spatial anti-aliasing) and fill rate.
Let's say for a moment that the product 3dfx will announce in the next few weeks is indeed producing a fill rate of around 900 Mtexels per second. In that case it would be the most powerful texel pusher we will see this fall.
Furthermore, their T-buffer technology will dazzle gamers running CURRENT games under DX6.1. No additional support will be needed to benefit by the difference of spatial anti-aliasing. Simulation and strategy gamers will be able to see the difference in F15, Falcon 4, Rogue Spear and other games NOW, without waiting for the new crop of DX7 games in the spring of 2000.
While I took the liberty of some speculation in my previous 3d hardware update, I am no longer at all sure which path will be the best this fall. We simply don't know enough about 3dfx coming product line to compare to the significant improvements we are hearing about from Nvidia and S3.
Which would you choose, given the alternative? A fill rate of 480 (perhaps more), 700, or 900 with full screen anti-aliasing NOW? Or would you buy for the future, with onboard geometry processing under DX7? Don't underestimate the power of full screen anti-aliasing. I remember my first look at a game at 640x480 with anti-aliasing and 32 bit color. I could have believed the resolution had doubled. No more jaggies will simply revolution games like Falcon 4, Panzer Elite, WW2 Fighters, EAW, B17 II and others. Anti-aliasing is that impressive, and 3dfx hopes to give it to us at resolutions up to 1024x768x32.
But while no more jaggies is nice, what about a huge improvement in frame rates? With the release of F15, F4 will likely be patched to a new graphics engine including DX7 support. That means every F4 nut out there might want the latest video hardware with onboard geometry processing. B17 II has been delayed to the new year partly because of the desire to add support for hardware T&L.
If 3dfx doesn't come up with incredible fill rates and the memory bandwidth to support it, or its own onboard geometry processor, I can't see how they could maintain their status in the industry. It's been a great race up to now, but we all know that 3dfx have seen what is coming, and we know they are determined to ride the cutting edge. We're all along for the ride this fall, and quite the trip it should be!
Join a discussion forum on this article by clicking HERE.
Copyright © 1997 - 2000 COMBATSIM.COM, INC. All Rights Reserved.
Last Updated September 7th, 1999