The State of 3D

October, 2004

by Josh Walrath

            Is this to say that the NV3x architecture was bad?  No, it was far from being bad.  When comparing the NV3x products to the competing products from XGI (Volari) or S3 (Deltachrome), the FX series trounces them.  Only when the FX series is held up to the light of the R3x0 architecture from ATI do the flaws come out.  NVIDIA lost considerable marketshare to ATI in the years since the R300 was released, but it did stay competitive throughout that time.  While being 2^nd best isn’t terrible, it certainly didn’t help shareholder confidence.

            Last year I characterized ATI’s R3x0 series as a “brute force” architecture.  I couldn’t have been more wrong.  The design team brought in from ArtX totally threw away the classic Radeon architecture, and learned from the mistakes of the R2x0 series of chips.  These forward looking thinkers decided on the “wide and shallow” philosophy of pipeline design.  The R300 had 8 parallel pixel pipelines that were floating point throughout (no internal integer support).  The R300 converted everything to FP24 internally, and so it had no need to integrate any type of internal integer support.  While this may have been overkill for DX 8.1 type content, it simplified and streamlined the overall architecture so that it could have outstanding performance in PS 2.0 applications.  The possible slight decrease in overall potential DX 8.1 performance was well worth the tradeoff for having class leading PS 2.0 performance.  The very fact that the R300 had 8 pixel pipelines made the DX 8.1 performance impact moot, as it had more than enough horsepower to run any current DX 8.1 game as fast, if not faster, than the competing NV3x products.

            ATI did not skimp on other features as well.  The anti-aliasing unit designed for the original R300 is still essentially the same for the current ATI products, and I personally consider it the best AA unit on the market.  ATI did take some shortcuts when implementing their anisotropic filtering, but this really isn’t a weakness for the architecture (though for quality, the GeForce FX series had the best looking anisotropic filtering implementation).  ATI did not feel that the quality improvements of higher precision aniso justified the transistor or performance expense.  NVIDIA obviously agrees with this, as their NV4x has a very similar performing and looking aniso output.

            While the R300 was not as flexible as the NV3x architecture, it was very focused on what was needed and followed the DX 9.0 spec for PS/VS 2.0 nearly to the letter.  It couldn’t do nearly the amount of instructions as the NV3x architecture, but in even today’s applications it is not held back at all.  In any comples design, tradeoffs are made between functionality, performance, and transistor budget.  With the R300 it appears that ATI hit the happy medium for each of the above mentioned aspects.

            The R3x0 is a very flexible architecture, and the basic design has been carried on through a wide range of products.  The latest X series of Radeon’s are still based off of the R3x0 architecture, but with some important changes and added flexibility (PS 2.0b).  ATI really did an amazing job with the R300, and the following products based off of that architecture.  It has allowed them to take a significant portion of the graphics market, as well as gain a lot of mindshare with general users and enthusiasts alike. 

The Current Situation

            This Spring brought an entirely new batch of chips and cards to the market, and it again changed the graphics landscape.  NVIDIA was first out of the gates with their next generation product, while ATI followed some weeks after.  These releases helped to level the gaming playing field, an area that ATI had controlled since the release of the Radeon 9700 Pro.

            NVIDIA released the GeForce 6800 series, which is based on the NV40 chip.  The NV40 is a major step forward for NVIDIA, and a large step away from the GeForce FX architecture.  NVIDIA adopted the “wide and shallow” chip design with a massively parallel 16 pixel pipeline product that was focused on floating point execution.  NVIDIA further pushed the design with the world’s first Shader Model 3.0 implementation.  While other areas were left fallow (it uses the same multi-sampling AA unit as the GeForce FX units- though with a different 4X RGMS algorithm), it was a massive leap forward in pixel pipeline design for NVIDIA.

            The basic pixel pipeline design features one full shader unit and one partial shader unit/texturing unit per pipeline.  This allows the NV4x pipeline to apply two shader ops per pixel, or a shader op and a texture op per pixel with each pass.  The design does not allow 2 shader ops and a texture op per pixel each pass.  With 16 such pipelines working together, a large amount of work can be done per clock.  The addition of SM 3.0 functionality (PS 3.0/VS 3.0) took up a large portion of the transistor budget for the NV4x series of chips, but it was one that NVIDIA felt was needed due to the competition that ATI brought to the market.  Through solid engineering, SM 3.0 functionality did not come at the expense of performance (unlike the relatively underperforming FX series of chips).  Also, the NV40 chips seem to be able to compete very well with ATI’s X800 series, even though the NV40 is clocked much, much lower.

Next: More NV4x

Copyright 1999-2004 PenStar Systems, LLC.