3dfx Interview with Peter Wicher

0 thumbs up

The VSA-100 - Voodoo4 & 5

Dispelling The Myth

A discussion with 3dfx Interactive

     

Thanks for stopping by folks. Tonight, we decided to have a little chat with our friends at 3dfx Interactive, Brian Burke P.R. Manager and Peter Wicher, Director of Product Marketing. It seems there has been some fairly serious debate around the net and in the industry about the architecture of 3dfx's next generation technology. Some believe that the next step for the Voodoo platform is not much more than an increase in clock speed with a few extra bells and whistles. We decided it was high time to get to the bottom of this debate and get some direct feedback from the people who are really "in the know". We asked 3dfx to let us have a peek under the hood of the Voodoo 4. Here is what they told us.




Dave:
OK, so Voodoo3? 5 layer metal, .25 micron technology? Refresh my memory here. What are the differences between the Voodoo3 and Voodoo4 Silicon? Are we talking larger data pipes and on chip memory for buffering? What are some of the physical differences between Voodoo3 and Voodoo4?

Peter:
To start with we're in a much nicer process technology. VSA-100 still uses a base 0.25um process geometry but we use six layers of metal instead of five to get better density and speed, and the actual transistor has a slightly shorter gate length and thinner gate oxide to make it go faster, too. The chip design does have a larger texture cache and the data paths are 32 bits rather than 16 (actually the rendering calculations are 40-bits wide in VSA-100 but the operands and results are stored as 32 bits). In terms of transistors we're looking at 14 million in VSA-100 versus 8 million in Voodoo3 so VSA-100's going to be physically larger.


Dave:
We've heard a lot about the addition of full 32bit color support, AGP4X and up to 64MB of memory. Is the VSA100 just a souped-up version of the Voodoo3? Can you tell us about the features that are exclusive to any of the next generation products based on the VSA100 and also those only available on the Voodoo5 Multi-Processor boards? People have said that the VSA-100 is just the result of a die shrink and a few new features that where left out of the Voodoo3 in an effort to get it to market in time? Any truth to that?


Peter:
Some VSA-100 features were easy to add to the Voodoo3 base, but others required significant redesign. Things like 32-bit rendering, 64MB memory support, and 4X AGP were not a great challenge. I suppose that we could have shipped this kind of "Voodoo3+" a long time ago but we didn't think that it was compelling for our core customers: we went a lot further. Real SLI is very hard to do, even for the guys who invented it. That's why four years after Voodoo Graphics no one else even has two-way SLI while we're up to 8-way on a single board. FXT1 and DXTC texture compression, combined with full 32-bit texture support and 2K x 2K texture size, was also a non-trivial feature addition and beyond what anyone else in the industry is doing. We also took time to trick out the basic 3D engine. VSA-100 is a full two pixel per clock architecture as opposed to Voodoo3's single pixel per clock, we improved the raster efficiency by about 20%, we increased the texture cache size, we added new texture combines and alpha blends, and we added guardband clipping which significantly offloads the CPU from doing 2D
clip operations to speed up applications.

The SLI boards, that is the Voodoo5 products, enable the full-scene antialiasing (AA) and T-Buffer effects and provide incredible pixel fill rates. Once you've seen a game with full-scene AA you'll never want to go back. Jaggies and scintillating pixels are totally annoying, they suck. One of the beauties of our AA is that it will work with any games now and in the future. Gamers who buy Voodoo5 will get an immediate, huge benefit on their favorite content. The insane fill rates of the SLI boards, like over 1.3 gigapixels on the Voodoo5 6000, will enable high frame rates at serious resolutions: no more knocking the res. down to 640x480 to survive a death match!

 


Follow up - 12/8/99

Davo:

We've heard a lot about the addition of full 32bit color support, AGP4X and up to 64MB of memory. Is the VSA100 just a souped-up version of the Voodoo3? Can you tell us about the features that are exclusive to any of the next generation products based on the VSA100 and also those only available on the Voodoo5 Multi-Processor boards? People have said that the VSA-100 is just the result of a die shrink and a few new features that where left out of the Voodoo3 in an effort to get it to market in time? Any truth to that?

Peter:

Some VSA-100 features were easy to add to the Voodoo3 base, but others required significant redesign. Things like 32-bit rendering, 64MB memory support, and 4X AGP were not a great challenge. I suppose that we could have shipped this kind of "Voodoo3+" a long time ago but we didn't think that it was compelling for our core customers: we went a lot further. Real SLI is very hard to do, even for the guys who invented it. That's why four years after Voodoo Graphics no one else even has two-way SLI while we're up to 8-way on a single board. FXT1 and DXTC texture compression, combined with full 32-bit texture support and 2K x 2K texture size, was also a non-trivial feature addition and beyond what anyone else in the industry is doing. We also took time to trick out the basic 3D engine. VSA-100 is a full two pixel per clock architecture as opposed to Voodoo3's single pixel per clock, we improved the raster efficiency by about 20%, we increased the texture cache size, we added new texture combines and alpha blends, and we added guardband clipping which significantly offloads the CPU from doing 2D
clip operations to speed up applications.


Peter:

Wow, my answer caused quite a stir, didn't it? Let me explain what I was trying to communicate about 4X AGP, 32-bit rendering, etc.

VSA-100 is a truly unique product. 3dfx customers want us to deliver products with features that are beyond what they can get from anyone else. If the VSA-100 only added 4x AGP, 32-bit rendering and 64MB memory support, to a Voodoo3, it would not have satisfied our core customers. To be compelling and to meet our customers' expectations, our VSA-100 engineering team put incredible energy into unique, technologically advanced features including real-time full-scene anti-aliasing, T-Buffer cinematic effects, and FXT1 and DXTC texture compression. A Voodoo3+ would have "de-focused" us from the goal of delivering the best 3D experience possible. PC Graphics is a very competitive market and an in-depth focus on a well-defined goal is essential to winning. End users are going to be very pleased when they fire up 3dfx's next generation of Voodoo
.


Sounds like good product planning to me. Guess what folks, people always find something to bitch about. John Q. Public is a tough customer. If you ask me, (not that I need an invite :) ) 3dfx may be a tad late but I think the decision to release true next generation product vrs. the "shrink and tweak" approach was the only way to do it. 'nuf said...

 



Dave:
I have heard rumor that some of these new features were already in the Voodoo1 chip, but just not "turned on", true?

Peter:
No, none of these features were in the original Voodoo Graphics. We're working with 14 million transistors in VSA-100 compared with about one million in Voodoo Graphics. It's like comparing a Z8 with a Model-A.



Dave:
Ahh, Multi-Processing how sweet it is when it is implemented correctly? So, SLI (scan line interleave) the old and the new, why don't we just plug a VSA100 in an AGP slot and another in a PCI slot? Personally, I am VERY happy to see you can do it all on one card. What key architectural differences exist in the VSA100 that allow you to achieve SLI this way? Does it work the same way SLI did on Voodoo2?


Peter:
The new SLI is different and much better than Voodoo2's. Voodoo2 used a comparatively simple analog scan line interleave scheme. VSA-100 enables a true, multi-chip digital communication and computation protocol. In the driver we can program how many scan lines each chip will rasterize, from one to 128 per chip. This lets us optimize for resolution and required fill rate. You've already noted another difference, that we can do SLI on a single card, which is a big win in saving slots and makes four and eight chip SLI physically practical. We can also sit on the AGP bus. The new SLI is a nice upgrade from Voodoo2's PCI based approach.

Dave:
I am drooling over one of those AAlchemy boards. Hey what's a mere $40K between friends, huh?! Can you hook me up with one of those bad boys?!!! I promise to be VERY gentle with it? :-)

Brian:
What do you have for collateral? Seriously, we may have some surprises up our sleeves for the Voodoo4 and Voodoo5 roll-out. You may be able to get your hands on one, but no promises yet.


Dave:
(whimper) Oh, sorry I was mopping the drool up there? Uh?. Hey thanks Bri! If my next born is male, the name is Brian and if female, well then she'll have to be named Briana! :-) So, what kind of performance increases can we expect from the Voodoo4 and Voodoo5 family of products. Do we have a "GeForce Killer" here or what?


Peter:
From what we know of the GeForce and it's upcoming 0.18um die shrink we're pretty comfortable. When end users compare Voodoo5's significantly higher fill rates, full-scene AA, and T-Buffer effects, not to mention our superior title compatibility, with what they will get from the 0.18um GeForce they're going to buy the Voodoo5. Looks better, runs faster: it's an easy decision.

Dave:
Won't the Voodoo4 4500 perform only marginally better Voodoo3 3000? Is it a similar graphics core running at the same speed? Why should anyone who already owns a Voodoo3 lay down $179 for a Voodoo4?

--->Page 2<---

Article Index:

Login or Register to Comment
Post a Comment