We have already told you that NV30, the upcoming next generation NVIDIA graphics chip, is not in the mass production yet. Moreover, it is still unclear if the guys already have the final GPU version or just another versions of the engineering samples. The curtain of mystery has been raised a little bit by our colleagues from ReactorCritical, who shared with us the confidential commentaries made by some NVIDIA employees about the words of their CEO.

It turned out that the first silicon of the notorious NV30 was made in May already, and they had about 3 months to complete the final chip version, so that it could start shipping in August already. According to NVIDIA’s initial plans, the cards on NV30 had to appear in October, which means that the mass production of the chips and the cards built on them could have started next month, if everything had followed the plans.

However, strictly speaking, the development of any contemporary chip is a very hard and thorough task, so that often different blocks of the chips are polished and finalized with the help of various emulators. Since at the very last stage, when there are about 5-7 parts to be combined together to build the final solid chip and since they communicate with one another via special scripts written “on the fly”, the bugs in the final silicons are simply inevitable. After the chips are actually built, the bugs are considered carefully to be then eliminated in the most efficient way. Bearing in mind that NV30 consists of over 120 million transistors, the bugs in the graphics processor may sometimes turn out pretty unexpected, which may even cause unplanned delays of the schedule.

Note that some analysts claim that the chips yields in case of 0.13micron TSMC process will make around 15%, which is not that much at all. So far, we have no idea what gives way to these estimates, as VIA has been using TSMC’s 0.13micron production lines for over a year now to produce their C3 processors.

All in all, the work on NV30 sticks to the plan very accurately. The chip may come out somewhat later because of certain production problems or bugs, say, some time in November-December, but this will hardly change anything in the market. All the chip features will be involved to the full extent only in 1.5-2 years at the earliest, which means that the graphics cards based on it will be targeted only for hardcore gamers and hardware enthusiasts, which are not so numerous. The only thing that upsets us a little bit is the fat that the delayed launch of the rumored NV30 will automatically postpone the other products, such as NV31 and NV34, planned for the next year.

But the most interesting piece of info on the new solutions is the anticipated performance of the NV30.

• In Quake3 1280x1024 with 4x FSAA Nv30 is expected to work 2.5 times faster than GeForce4 Ti4600.
• In 3DMark2001 SE NV30 will perform 3 times faster than the predecessors.
• Taking into account NV30 architecture, it will prove 3.5 times faster than GeForce4 Ti4600 in Quake3.
• The effective fillrate of NV30 in case of 2x anisotropic filtering will be 2.7 times higher than by the today’s flagman.
• In case of pixel and vertex shaders, we expect this solution to show really great performance increase. In particular, pixel shaders will be processed 4 times faster (it is most likely to be about DirectX 8.1 shaders), and the GPU processing speed will grow up to 200 million polygons per second.

This incredible performance growth can be explained firstly by several reasons. Firstly, the number of pipelines grew up to 8 (with 2 TMUs each), and secondly, the effective memory bandwidth reached 48GB/sec. The memory in this case will work at 1GHz frequency and will communicate with the chip via 256bit bus.

Speaking about the architectural peculiarities of the NV30 chip, we should say that this Cinematic Shading GPU boasts a number of indisputable advantages over the existing architectures. Namely, it supports complex shaders programmed on NVIDIA Cg language with a great lot of syntactical constructions, branches, cycles, etc. I would also like to say that NVIDIA reports about some continuous shaders, which is really hard to believe, especially as it is about real time calculations. This implies that the chip should feature special buffers storing intermediate rendering results (if necessary). All buffers are of limited size, that is why it seems too early to talk about these innovations before the chip is actually launched.