Jen-Hsun?s comments at the start of the recent NVIDIA Editors Day suggested that they had taken onboard the principals that ATI set forth with R300 of going for a very parallel architecture. As you look beyond the wide pipelined nature of NV40 and begin to look at little more at the pixel shader composition and the various quality options available you begin to see that this is not all NVIDIA have adopted. NV40 is not a particularly revolutionary architecture, but very evolutionary from many of the principals ATI delivered on some 20 months ago, combined with some of the better elements of NVIDIA?s previous architectures.
Such an approach is no bad thing as we can see it has delivered for ATI. However, NVIDIA have taken many of these principals to the next step and evolved them further. The ?Superscalar? nature of the pixel shaders should provide an extra level of flexibility that NVIDIA?s driver compiler should be able to take hold of and increase the shader performance. Although NVIDIA have opted to stick with an ALU that?s also used for texture address processing, this is probably no bad thing either because as shader lengths increase the ration of texture accessed to ALU use will drop, reducing the need for a dedicated texture address processing unit.
Of course, not only do we have the evolutionary steps, we also have some great introductions as well. Higher speed FP32 rendering throughout the pipeline, the introduction of Shader Model 3.0 in hardware and the Higher Dynamic Range Rendering being key highlights that NV4x brings for the first time. As ever, the likelihood of many of these being adopted on a wide scale basis in games in the NV4x lifetime are unlikely, its good to have the options there for those that do and it also puts the capabilities in the hands of developers for the future.
With regards the the performance of GeForce 6800 Ultra we can see that there are massively impressive gains over the FX boards in many cases. We really have reached the point where far too many titles are completely CPU bound, and now high quality options such as FSAA and Anisotropic filtering come for free at much higher levels. However, given this it's still odd that NVIDIA haven't implemented a more effective mode of FSAA beyond 4X FSAA - the rotated grid sampling pattern for 4X FSAA is a welcome addition to NVIDIA's parts though.
Thanks to the new shader pipelines, and the sheer number of them, the shader performance takes a very large performance leap over the FX series. The new Pixel Shader pipelines see very large gains in performance with short length shaders; however, do note that while the performance of longer shaders is still much greater on the 6800 Ultra it does appear that the register space issues aren't entirely removed as witnessed by the performance increases in FP16 performance on longer shaders over FP32.
Naturally, performance will come at a cost and in the case of NV40 one of the main costs is power. While NVIDIA recommends a 480W PSU the test rig here ran fine with GeForce 6800 Ultra on a 450W power supply, though this isn't exactly overburdened with extra components - regardless, it looks like many that wish to run a 6800 Ultra will require a new PSU as well. Heat may be another factor - the test system used here is an open system, so cooling is an issue, yet for those with a closed box may wish to ensure there is plenty of air circulation in there as well.
NV40 appears to be a hugely ambitious piece of engineering and it has certainly seen the benefits in terms of performance in relation to its predecessors. The new technology implementations of Shader Model 3.0 and Higher Dynamic Range rendering are also welcome additions to the technology available to developers, giving them a wider range of flexibilities for next generation gaming titles. When all is said and done, this is a very impressive, and welcome, leap over the FX series.
Del med dine venner