Preview of 3Dfx Voodoo Banshee, S3 Savage3D and NVIDIA RIVA TNT

Page 1 of 7:

Introduction

The late summer and fall 1998 will bring us a bunch of new 3D graphics chips, most of them already announced for quite a while. The expectations are high, since the chip announcements were sounding miraculous in many cases. This article is supposed to give a preview of some of the new chips. Some of them are still in early stages, so that there should be room for performance improvements. Before I get into the actual review data, I would like to quickly go through the chips that are supposed to be available soon, discussing the facts, hype and rumors about them.

Banshee is 3Dfx' first 2D/3D solution, unlike the Voodoo Rush it doesn't require an additional 2D part. 3Dfx was putting a lot of effort into the 2D hardware, offering 'all Windows GDI functions in hardware', thus saving a software layer between the driver and the graphics hardware. This is supposed to make it the fastest 2D chip available. It offers an integrated 250 MHz RAMDAC, to enable good quality 2D. The 128 bit 2D core is also supposed to make Banshee the fastest DOS game platform, important for people who still like re-playing System Shock (I love it) or Privateer II. Banshee also comes with 'full DVD support', enabling 'hardware assist' for software DVD and a 'full' VMI interface for single slot hardware DVD.

Latest Videos FromTom's Hardware

Watch full video here:

The most important thing about Banshee is of course the 3D core. Here 3Dfx was being smart and simple at the same time. Banshee has an only slightly different 3D interface as Voodoo2 and can so not only run the Direct3D software as all its competitors, but also all the numerous Glide titles programmed for Voodoo and Voodoo2 as soon as Glide 3.0 is available with it. This will be the greatest strength of Banshee, making it look considerably more appealing than all its competitors with their young proprietary 3D engines, which lack any special software support.

The main difference between Banshee's 3D engine and Voodoo2 is the missing 2nd texture unit of Banshee. This second 'TMU' enables Voodoo2 to render complicated scenes using both texture units in parallel and thus faster than if it's done sequentially. Quake 2 is a typical example, it uses at least two 'passes', which can be done at the same time with Voodoo2. Banshee will not be able to do that. On the other hand the raw fill rate of Banshee is higher than Voodoo2's. Whilst Voodoo2 runs at 90 MHz and thus offering 90 MPixels/s, Banshee will do 100/125 MPixels/s. Games that don't use 'multi-pass texturing' will so run slightly faster on Banshee than on Voodoo2. Those games are right now in the vast majority, but this will change in the nearer future.

After looking at Banshee's specs, it seems obvious that Banshee will be a great success. It runs the majority of the current (single-pass) Direct3D games faster than a single Voodoo2 card, it runs all the Glide games (e.g. Unreal) as well and multi-texturing games slower, but not that much slower than single Voodoo2. It offers a very good 2D engine and it comes at an attractive enough price. This is exactly what the market was waiting for. Scott Sellers opinion that game developers need a reliable platform which has a decent longevity, will show to be absolutely right, particularly with Banshee. The Glide support makes it unrivaled in the market of 2D/3D chips.

G200 was already discussed here before and the performance of roughly about 85% of a single Voodoo2 chip showed that G200 will be an attractive product. Different to Voodoo2, the G200 chip offers 32 bit rendering, 32 bit Z-buffer and several other quality enhancements, which make it a very attractive product for people who want fast 3D without quality trade-offs. The 2D performance is good as what we are used to from Matrox. G200 is now already available as 'Millennium G200', including a 250 MHz RAMDAC, targeted to the business market and soon there'll be the 'Mystique G200', equipped with a 230 MHz RAMDAC, but several game bundles. The OpenGL ICD is unfortunately still not finished yet, but the current D3D wrapper shows that the G200 will run Quake 2 fast and at high quality too. Matrox' name, famous for speed and quality will be required to sell G200 against the upcoming Banshee products. Although G200's quality is better in 3D and possibly even 2D, the 3D performance of Banshee combined with its Glide support make it look more attractive.

The youngest chip in this discussion is TNT and it seems farest from completion as well. The specs of TNT blew away everybody when NVIDIA announced it, 250 MPixels/s, 8 million triangles/s are amazing numbers. However, the requirements for reaching those numbers are very high too. First of all TNT will have 7 million transistors, which is as many as a Pentium II core. This makes the chip big and expensive as well as not exactly easy to handle. Secondly this chip will have serious thermal problems when manufactured at 0.35 micron and running at 125 MHz. NVIDIA admits that the shrink to 0.25 micron will not be before the end of this year and that 125 MHz will not be possible with 0.35 micron. Rumors say it will only be 100 MHz, reducing the peak fill rate already down to 200 MPixels/s. Another rumor sounds even more scary. TNT is as Voodoo2 equipped with two texture units, thus enabling dual-pass as well. Each texture unit is able to supply 125 or rather 100 MPixels/s. Multi-pass texturing will get more and more popular in future games, so that TNT is as attractive for future gaming software as Voodoo2 is. Now the latest news from DirectX6 Melt-Down is that TNT may only ship with ONE texture unit, reducing the fill rate to 100 MPixels/s at 100 MHz and thus making TNT a lot less attractive. I really hope that this is not true, since this would jeopardize TNT performance considerably.

Now let's get back to the facts, who knows how much truth is about those rumors. The TNT will have a very good 2D engine and the 3D engine should be extremely powerful too. Whilst Voodoo2 can deliver 90 MPixels/s fill rate which each of its 'Texelfx2' units, the RIVA can deliver 100-125 MPixels/s which each of its 'texel pipelines'. This should make it faster, but not that much faster than Voodoo2. It is unlikely that it will surpass SLI Voodoo2, but on the other hand it can offer higher resolutions at still remarkably high frame rates. TNT can render 32 bit color and has a 24 bit deep Z-buffer, which is supposed to be just as exact as the 32 bit Z-buffer of G200, thus enabling higher 3D quality than current products from 3Dfx. TNT has the chance to be a better and more cost effective product than Voodoo2, so the question is only if NVIDIA can indeed deliver this high performance of TNT, if it will be at an affordable price and particularly when this will be. One thing is for sure though, TNT will not be compatible to Glide, which accounts for one serious disadvantage over Voodoo2 and its successors.

After holding the crown of the mainstream graphic chip business for a long time, S3 almost disappeared in 1997. Products with S3 chips were a clear no-go unless you had to get something really cheap. Remember the words of Alex 'Sharky' Ross, calling the Virge chips '3D decelerators'. Diamond used to be S3's number one OEM and last year they told me that they were happy cutting the connection to S3 finally. The only area where S3 was still having a good name was notebook graphic chips.

Now S3 wants to get back into the business and they've got a hard time. S3 made itself such a bad name that nobody really wants to believe that the new Savage3D is indeed a decent performer. Too loud were the words of S3 themselves, claiming 'Voodoo2 performance at a much lower price'. This test shall show if there's some truth about 'Voodoo2 performance'. So far many tried to reach or even surpass 3Dfx products, alas so far nobody was able to.

The Savage3D architecture doesn't sound bad at all. It has a full setup engine and a texture cache as nowadays common in all good 3D chips, it's equipped with two pipelines, one for rendering, one dedicated for texture processing and it can do texture compression, keeping the texture memory bandwidth low. You will also find AGP 2x support, not that this is a really important technical feature, but Intel made it to a very important marketing feature instead, so a 3D chip's better going to have it. The Savage3D is able to do tri-linear filtering in one cycle, offers 5 million triangles/s and has a fill rate of 125 MPixels/s,it can do 32 bit color rendering as well. As the G200 it is not equipped with one 128 bit memory interface, but two 64 bit ones. The Z-buffer can be 16 or 24 bit deep. The Savage3D supports only up to 8 MB local memory, all the above chips support 16 MB.

So far the feedback from OEMs was mixed, the chip seems to be as expensive as Banshee but without the name, the Glide support and seemingly with less performance as well. I personally think that if the i740 with its average 3D performance could become successful, although it was released after Voodoo2, then the Savage3D should have at least as much of a chance. It is definitely in the new high performance area, the i740 certainly isn't.

Well, what shall I say about Videologic and the PVRSG? After making a lot of noise at CeBIT, which admittedly caught my attention, Videologic has seemingly gone into 'sleep-mode' regarding PVRSG. Rumors say that Videologic is overly busy with their SEGA deal and have no resources for PVRSG. Another rumor says that the UK office has been or is about to be closed down. Uops! Whatever Videologic is doing, it doesn't seem to be much about the 'highly anticipated' PVRSG and if they think that they can treat their followership in the PC area with this kind of disrespect, by preferring to do a gaming console deal with SEGA, then PVRSG does certainly not deserve a plug from Tom's Hardware Guide either at this time.

First of all I'd like to make clear that none of the three cards is in final state as yet. Banshee is pre-release, running the alpha2 driver, Savage3D is in pre-release, running the 07 beta driver, TNT is in very early state, running at only 70 MHz and using an alpha driver as well. Thus the TNT will certainly not show final performance at all, Savage3D will hopefully have more stable drivers soon and Banshee will probably get faster as well. Whilst I had half a day for Banshee testing and a lot of time for Savage3D, the TNT was here only for about 2.5 hours. This is why I could test TNT with only one CPU. I could have done and thought of quite a few more tests for Banshee, but time wouldn't allow this. In the next week I should at least be able to give you results of a 16 MB Banshee board and a more thorough testing as well. S3 prefers positive press, which is why they wouldn't supply me with a board on their own. As you can see that doesn't mean I wouldn't get one, but this way I may not have the latest drivers and I was missing OpenGL drivers for running Quake2.

8 MB board, driver 4.10.01.4000-6.00.07A
display buffer swaps were not synchronized with refresh rate, as visible when looking at Forsaken frame rate of 110/s. However, the refresh rate has got an impact on frame rate, which is very odd. It seems as if the driver can recognize some benchmarks and switch on/off "VSync".
Unfortunately I did not have an OpenGL driver for Savage3D, so that I could not run Quake 2.

8 MB board, driver 4.10.01.4000-6.00.07A
display buffer swaps were not synchronized with refresh rate, as visible when looking at Forsaken frame rate of 110/s. However, the refresh rate has got an impact on frame rate, which is very odd. It seems as if the driver can recognize some benchmarks and switch on/off "VSync".
Unfortunately I did not have an OpenGL driver for Savage3D, so that I could not run Quake 2.