   |
|||
|
|||
|
:
News
: Archives
|
|
The State of 3D |
|
|
Economies of 3D |
|||
|
April, 2006 |
|||
|
by Josh Walrath |
|||
|
The NV4x series were large chips and the initial manufacturing of the chips was fairly poor on IBM’s 130 nm FSG process. These problems eventually cleared up and NVIDIA sold a lot of cards based on the technology. Their more successful chips were the 6600 and 6200 series, and these were based on the smaller TSMC 110 nm process. Still, the GeForce 6800 GT and Ultra models sold very well, and the introduction of SLI added another option and value added feature. NVIDIA certainly looked to be pushing the envelope again with this generation of product. Typically after a major architectural change there is implemented a refresh six to eight months after the initial product. NVIDIA skipped this while ATI did not. This left many scratching their head wondering what was going on. In the late Spring of 2005 speculation was rife about what NVIDIA would do to counter ATI’s overwhelming looking R520. The R520 had been taped out in late November of 2004, and its release in late Spring of 2005 seemed imminent. Many were expecting a major new architecture from NVIDIA that would encompass new features and performance. The GeForce 7 series of cards were eventually released in June of 2005, and there many mixed emotions about this new product. This was more or less a heavy refresh of the NV40, which featured more pixel shader and vertex shader pipelines, greater per clock efficiency in the pixel pipelines, a small increase in clock speed, enhanced ability to render at resolutions above 1600 x 1200, and the introduction of transparency AA. This was not a huge stretch of the NV4x architecture, but the success of this architecture was guaranteed by the serendipitous delay of ATI’s R520 and family. This was an architecturally conservative move by NVIDIA, and one that seemed justified by the issues that ATI was having with its new design on TSMC’s new 90 nm Low-K process. NVIDIA was slower than the competition in implementing a large design on the 90 nm process, but in hindsight this appears to be a good decision. This is in direct contrast to NVIDIA’s earlier history where they were typically one of the first of the fables semi’s to utilize new process nodes. NVIDIA was hit pretty hard with the whole GeForce FX, and perhaps it was this experience that gives NVIDIA pause about jumping headfirst into a new process node. NVIDIA’s first 90 nm part was the integrated GeForce 6100/6150 chips for motherboards, and that was a far less complex part than what ATI was trying to put out. It wasn’t until January of this year that NVIDIA officially released a standalone desktop part with the GeForce 7300 based on 90 nm Low-K. Until the release of the 7300, many were expecting NVIDIA to introduce a more forward looking architecture with more pixel shader pipelines, improved texturing quality, and redesigned ROPS that can handle simultaneous HDR and multi-sampling AA. When NVIDIA did release its 7600 and 7900 series of parts, people were a trifle disappointed by the omission of any of the advanced features that many were hoping for. This was offset though by the fascination about what NVIDIA ended up doing with the architecture. NVIDIA achieved a new level of conservatism with the G71 and G73 parts. Not only did NVIDIA retain the complete featureset of the GeForce 7 series, but it actually cut down the transistor counts of each of the chips by a pretty significant margin. Not only did they cut down the count, but they significantly improved per transistor performance. Because of the transistor performance NVIDIA saw with the switch from 110 nm FSG to 90 nm Low-K, they discovered they can still achieve their clock goals while removing pipeline stages, thereby cutting down the overall transistor count. The original G70 had over 302 million transistors while the G71 now features 278 million. The G73 has a count of around 178 million while its functionally identical predecessor, the NV42, had a count of 199 million. This was a very conservative move for NVIDIA by not only retaining the basic architecture, but in fact lowering the transistor count while keeping overall per clock performance. NVIDIA is actually quite pleased with this, as they have kept their products very competitive in terms of features and performance, yet have lowered the cost of production for their entire lineup by a great deal. This is quite different from the NVIDIA of old, one which said “die size be damned” and continually pushed the feature envelope. ATI has had a very different outlook through their entire history of 3D products. At the time of the Voodoo Graphics and Riva 128, ATI added basic 3D features to their Rage series of products. These were never great 3D performers, and were actually only a small step above the infamous S3 Virge series of chips. ATI’s first truly successful 3D chip was the Rage 128, and it had performance that was slightly faster than the competing Riva TnT, but unfortunately for ATI the Rage 128 was released just before NVIDIA unveiled the TnT 2. ATI always seemed about 8 months behind NVIDIA with their products, but typically these products were a bit more progressive than NVIDIA’s. The Rage 128 was more tuned to run in 32 bit mode, while NVIDIA seemed to perform much better when just using 16 bit. The Rage 128 Pro was faster than the TnT 2, but it was very late to market and never seemed to come out in great numbers.
If you have found this article interesting or a great help, please donate to this site.
Copyright 1999-2006 PenStar Systems, LLC. |
|||
