X-bit labs

ATI RADEON 9600 XT vs. NVIDIA GeForce FX5700 Ultra (page 7)

Category: Video

by Alexey Stepin

[ 11/03/2003 | 10:33 PM ]

NVIDIA seems to have realized clearly all the pros and cons of cheating now. We do not expect the company to perform any more of its hardcore optimizations, such as those for 3DMark03 found by FutureMark earlier this year. Instead, NVIDIA now claims that its excellent software optimizes virtually every single shader program to run better on GeForce FX-series GPUs. Since no one can deny NVIDIA’s right to optimize the code run by its hardware, the company will continue the development of its software optimization layer called “Instruction Optimizer”. In case it brings some tangible speed boost, other leading graphics companies, particularly ATI Technologies may also follow the suit. There are a number of implications, however:

• It is not effective from any point of view to develop software that would provide just a slight performance boost, but would take a year to be developed. But NVIDIA seems to be in a position to afford the cost both in terms of time and money.
• If general optimizations of DirectX and OpenGL shaders are adopted by both leading GPU companies, NVIDIA and ATI, and if it does cost quite a lot of money to make the “Instruction Optimizer”, the top-tier graphics companies will have another huge benefit over newcomers, such as XGI. Moreover, this advantage is very likely to close this market for new companies altogether.
• The importance of DirectX or OpenGL lowers, as IHVs implement software that optimizes shaders’ code.
• It is extremely positive for NVIDIA’s customers that the company really does something to boost shader performance on the GeForce FX hardware. We should applaud NVIDIA for this intention.

The biggest issue with benchmarks for graphics cards testing is the fact that they measure performance in one and only environment. Sometimes, even different demos can reveal totally different and contradicting results in the same game. Moreover, opposite results may emerge even if we test one particular demo in different versions of the game. Needless to say, even gaming benchmarks are not precise performance indicators for games. So, it is rather strange that we expect synthetic benchmarks to show us anything clearly. 

Well, let us sum everything up now.

1. NVIDIA GeForce FX is a good performer in a vast majority of today’s games, but unless NVIDIA convinces game developers to degrade image quality dramatically, the whole GeForce FX family will be relatively slow in tomorrow’s games.
2. Cheats or other tricks in drivers aimed at improving performance are not a panacea here because they require tremendous and cost-ineffective work of NVIDIA’s software team. No company will ever afford to improve performance in absolutely all games by adjusting the software. In order to really improve the speed, IHVs have to polish off the hardware.
3. Even if the NV40 family products are excellent performers, end-users may have suspicions that the high performance comes from cheats while the real power is lower. This attitude of the users is too bad for NVIDIA, even worse than the obvious fact that the GeForce FX family is slower than the R300-family.
4. Most benchmarks just show us how fast this or that GPU is in this given benchmark. The benchmark never provides a thorough and exact picture of what a GPU is, unless it is in-depth and synthetic. Nevertheless, everyone will continue to use benchmarks to determine the best graphics cards or processors and make conclusions based on those results, even keeping in mind that those results are abstract. In order to avoid publication of incorrect results because of cheating in popular benchmarks and even in popular demos, the press will have to continue changing the list of the demos and benchmarks.