Sure - multi-chip can be cheaper than single chip whenever a single chip gets very large. Chip yields are such that they drop off very quickly after a certain die size, so two small chips can be cheaper than one large chip. And if you are starving for memory bandwidth, two chips allows you to have more memory bandwidth without increasing the pin count per chip, which can get very expensive after a certain point. That is why you don't see single chip 256-bit memory interfaces in the consumer market.
Reverend: In the same vein, wouldn't a multi-chip solution simply be less profitable? Do you see 3dfx dumping this multi-chip approach in the future? Or will its inherent scalability (for a variety of products based on the same generation, much like what we see via the Voodoo4s, Voodoo5 5500 and Voodoo5 6000) be as much an important factor in design?
If multiple chips are cheaper or even slightly more expensive than a single chip, but offer higher performance, I see no reason why they should be less profitable. If you can reach higher levels of performance with multiple chips, you can usually extract higher margins in these products. You will see scalable solutions from 3dfx in the future. I'm not saying that every product will be scalable, just that we will continue to move forward with product scalability in mind.
Dave: Deferred rendering is well known for the ability to display scenes very efficiently. An example of this is Gigapixel's design and technology. However, in the next couple of years these architectures are going to run into the problem of having to bin large amounts of much geometry data. With 3dfx (presumably) going to this type of architecture, that might bring concerns to the table. Are you confident that 3dfx will be able to get around this issue, or have you already found solutions? Can you elaborate?
We are confident that the GigaPixel design is the best solution even when dealing with the large amounts of geometry data that we expect to see in the future.
Wavey: There has been some discussion on the web that the method of FSAA employed by the T-buffer in the Voodoo5 is not Super-Sampling (i.e. one texture sample per sub-pixel) but Multi-Sampling (i.e. a single texture sample for the coverage mask of all the subsamples) - which is it, given that Scott Sellers has already told us that Multi-Sampling can work with the T-Buffer? If it is Super-Sampling, does that not mean that Voodoo5 has to resort to multi-pass rendering in multitexture applications, given that it only has one TMU per pipe?
Voodoo5 performs one texture sample per sub-pixel due to the nature of its architecture, which is a multi-chip architecture. So in that sense - Voodoo5 FSAA is like super-sampling. But in other ways, e.g. FSAA being a transparent operation to the application, it is closer to multi-sampling. Either name you call it, the Voodoo5 T-buffer does not require multi-pass rendering in multitexture applications. The T-Buffer concept works with both super-sampling and multi-sampling architectures.
Wavey: Scott Sellers recently hinted at improved FSAA performance on 3dfx's next generation product (the one widely known as Rampage) in his interview with me at ECTS. Now, we know VSA-100 Super Samples into multiple buffers using a jittered vertex offset; we also know that GigaPixel Multisamples with a pixel pipeline that automatically generates 4 subsamples - given that we've heard the GP tech will not be used in Rampage (because it's too far down the line) where can this extra FSAA performance in Rampage be achieved?
No comment at this time.
Reverend: Would having a dedicated pipeline just for FSAA be feasible to improve performance?
Not that I can see - FSAA requires rendering multiple subsamples which pretty much occupies all existing resources. I don't see how you would dedicate just one pipeline to FSAA and not the others.
