GeForce 2 GTS Preview
By Björn Tidal and Scott Sherman
Updated 26 April 2000

Introduction

In 1996 something big happened in games. A company called 3dfx released a new type of add-on card called the Voodoo1. No one who is serious about gaming can forget the sensation as you looked at the first 3D accelerated screenshots from Tombraider. Sure, we had seen a few types of 3D accelerators before that but neither the Creative 3D VL Blaster nor the Diamond Edge was a success as they were before their time. The Voodoo1 on the other hand offered something completely new. Gaming at high (!) framerates (you could actually go over 25 fps in some games) and new standards in image-quality was finally here. Right after the launch of the Voodoo1 came the first Rendition card, the V1000 that had even better image quality even though it sometimes lagged a bit behind in performance.  

I would call this the golden age of 3D accelerators. Everything was new and exciting. Games with 3D support were rare. After this golden age things went downhill. New cards came but the main focus were speed. “Framerate is king” became the battle-cry as new chips entered the arena. Suddenly you couldn’t play a game if you didn’t have a framerate of 60 fps. Things became boring for this writer.

It all changed last autumn. More specifically when NVidia released their GeForce chip. The GeForce was the first in a new breed of cards. Instead of just concentrating on framerates and fillrate NVidia decided to take a stab at hardware T&L. NVidia must have know then that the games using the hardware T&L support wouldn’t have time to come out during the brief 6 month life of the GeForce. However NVidia was in a chicken/egg situation. Developers wouldn’t start to support HW T&L in their games unless there were hardware support. NVidia therefore deserves some applauds for putting themselves into the frontline.

One big reason (not forgetting it is a great chipset!) the GeForce went to become the best card in the market of course is because of 3dfx’s delays. The V4/5 was supposed to go up against the GeForce and judging from the first previews, it would have made a good competitor to the GeForce. But alas, we all know what happened. Delays delays delays. 

On the 21 April Scott visited NVidia for an introduction on their latest product – the GeForce 2 GTS. Scott also received a reference board for testing and this preview will include some preliminary benchmarks. More extensive testing will be done and the scores added during the coming weeks.

The GeForce 2 GTS

One thing NVidia stressed at the event was that the GeForce 2 GTS (which I probably will call GF2 in the rest of the preview) is not just a die-shrink’ed GeForce. It’s a bigger step between the GeForce and the GeForce 2 GTS than it is between the TNT and the TNT2. This means that NVidia actually have deviated to their normal plan of operations which so far has meant that they release a refresh every 6 month of their products. This is NOT just a refresh of the GeForce.

GTS in GeForce 2 GTS stands for GigaTexelShader. We will explain this a bit further down in this preview.

Let’s look at the specs of this baby:

.18 micron design – The die-shrink (the GeForce used .22 micron)  allows the GeForce 2 GTS to run at a core speed of 200 MHz. Not bad at all. The memory that is used is 333 MHz DDR-Ram (ok, I know it’s actually 166 MHz). No SDRAM versions will be offered. Considering that the GeForce itself already was limited when using SDRAM this shouldn’t' come as a surprise. The issue regarding memory bandwidth still remains even though NVidia says the bandwidth now is 5.3+ GB/s.

8 Texels per clock – The "old" GeForce had a QuadPipe rendering engine which could render four pixels per clock with one texture each. At 120 Mhz this gave a fillrate of 480 Mpps fillrate and 480 MTexels/s. The GeForce 2 GTS though has something NVidia's marketing guys has called HyperTexel pipelines. In fact- it has 4 of them. Each of these pipelines can process one 32-bit pixel per clock for a total of four pixels per clock. Each of the pipelines also can draw 2 textures on each pixel in one pass. This means that at 200 MHz we get a fillrate of 1.6 GigaTexels/second or 800 MPixels/Second.

Improved T&L – It would be strange if NVidia didn’t try to improve the T&L engine also. Well, don’t worry. The GF2 now can reach up to 25 Million Polygons/sec compared to the GeForce which could do 15 Mill Pol./sec. Before any of you besserwissers out there start to object – we all know that this is theoretical. The GTS also can do 8 hardware light per Traingle (not per scene). 

Full hardware FSAA – Yep. You eyes are seeing correctly. The main feature 3dfx is touting is now available on the GF2. You will be able to select from 2x up to 16x FSAA. I will talk about FSAA a bit later in this preview though as things aren’t exactly as they seem.

Texel Shader – Not wanting to just enhance the features in the GeForce the GF2 also has some cool new stuff. One is the NVidia Shading Rasterizer (NSR). This gives the GPU 7 Pixel operations in a single pass on each of 4 Pixel pipes. 

Most of these operations are common but the ones that caught my eye were the per-pixel lightning, per-pixel bumpmapping and per-pixel diffuse lightning.  I will gladly admit that I'm out of my deepths here so I won't try to describe what everything is. NVidia will most certainly have some white-paper up when you read this talking about this. The cool thing is of course that a developer can perform these 7 operations in one pass. Of course this is something that must be coded but let's hope the developers catch on and start taking advantage of it. NVidia though seems confident they will be able to get support and did show us some demos and screenshots from upcoming games. They also (as expected ;) had some quotes from developers talking how good this feature is. Let’s hope the content arrives soon.

Texture compression – The GF 2 has full support for texture compression both in D3D and OpenGL via DXTC and S3TC (from the license agreement with S3). This is an important feature due to memory bandwidth limitations.

Cool video features – The GF2 has full HDVP processor for video and DVD with Video bandwidth of 150MB/s (compared to 25 Mb/s for the GeForce). The GF2 also supports Timshifting and Digital VCR.

According to NVidia we will see boards as soon as a few days into May on store shelves. The prices we have received lies around 350$ for a 32 Mb card and up to 450$ for a 64 Mb card.  

Te GeForce 2 of course still has a digital output for LCD screens. This time it supports a resolution up to 1280x1024 which should be enough for *most* people. I know I have read about screens with higher res. somewhere but I guess for the average Joe this is enough support.

FSAA

Before we move to the benchmarks I would like to take some time talking about FSAA (Full Screen Anti Aliasing).

NVidia and 3dfx has decided to go separate ways. While 3dfx decided that FSAA was the most important feature NVidia decided that adding hardware T&L was the most important feature. Very carelessly spoken you can say that 3dfx went with “removing” some details while NVidia decided to add more details.

So which was is the correct one? And doesn’t the GeForce 2 GTS actually support FSAA which makes this whole discussion moot? Well, not exactly.

First of all – why would we want to have FSAA? Those of us who has used the Rendition cards actually have some experience with AA, even though I think it was edge-AA and not FSAA. In Vquake we could turn on different levels of AA and it did look good. The star performance though is in Indycar Racing where all the “jaggies” disappeared.

This is just an example pic. The triangle on the left is not AA’ed while the one to the right is.  The idea with FSAA is to remove artefacts and to make the image look “better” and “more real”. While a game like Quake 3 not might be the best game to use it in I know that flight-sim buffs and those who love car racing games especially look forward to FSAA. One big reason is that FSAA can help removing those annoying “shimmering” effects you get when you have faint lines who disappear and appear since they are only 1 pixel wide. With FSAA they will stay present on the screen all the time-

The biggest problem with FSAA if the simple fact that it needs a lot of fillrate. A lot(!). I expect to be corrected if I’m wrong here but AFAIK you will half the fillrate at 2x FSAA and divide it in 4 when using 4x FSAA. Some nasty rumours suggest that 3dfx needed something to use the huge fillrate of the V5 and therefore added FSAA. Regardless of the reason they incorporated FSAA in the V5’s and this is the main feature of the V5.

NVidia instead went on a different route. Instead of removing artefacts and “blur” the image (I use blur in a positive sense since the final result is intended to make it look better and not just blurrier) they decided to give the developers the chance to add more details to their games by supporting Transform&Lightning in hardware and thus allowing developers to use more polygons. This actually can add to the aliasing on an image but NVidia’s philosphofy is that it’s better to add details than remove it.

NVidia actually does support FSAA in the GF2. However it’s important to realize that the way NVidia employs FSAA is a bit different than 3dfx does. The 3dfx way is a bit better and gives a bit better quality than the NVidia way. If it is enough quality difference still needs to be answered.  

It's actually a little strange situation - for the first time 3dfx is pushing image quality over framerate while NVidia are pushing framerates.

I asked Scott to turn on FSAA in some of the games he played and tell me what he thought. One of the games he tried was Falcon 4 – a flight-sim that should benefit from FSAA. In Scott’s case though he clearly preferred to turn up the resolution a lot instead of turning the resolution down and turning on FSAA.  I get the impression that it really depends on the person when it comes to preferring turning up the resolution or using FSAA.  Since it’s beta drivers yet we want to leave FSAA until we get final drivers before we discuss quality. There might even be a possibility that it isn't working properly yet on Scott's machine so we don't want to draw any conclusions on this yet.

Okidoki - on to the benchmarks