GeForce 8: Scalability is Everything, Meet the GeForce 8600
For years Nvidia has been developing graphics architectures which allow easy scaling upwards and downwards, as after creation of a flagship graphics processor it is possible to design lower-end chips without many problems based on the building blocks of the higher-end part. The GeForce 8-series is not an exception and, in fact, the unified shader architecture allows taking the approach one step forward.
The GeForce 8800 GTX consists of 8 “blocks”, each of which features 2 shader processors (one shader processor sports 8 stream processors or ALUs), 4 texture modules and a shared L1 cache. Besides, each block has an array of general-purpose registers, L2 cache and 64-bit memory controller, which can be accessed by any other “blocks”. Such approach allows Nvidia to easily disable defective parts of a particular graphics processor to create cut-down solutions, such as the GeForce 8800 GTS, based on the same chip, or to develop new processors without major changes when it comes to the architecture.
The newly released GeForce 8600-series graphics chips features two “blocks” with 32 stream processors (which work at 1.45GHz), 16 texture modules, 8 raster operation units (ROPs) and 128-bit memory controller. In fact, configuration of the 8600-series seem be rather strange for Nvidia: back in the days the company used to offer performance-mainstream GPUs with half the number of execution units compared to high-end/premium parts (e.g., the GeForce 6600 featured 8 pixel processors, whereas the GeForce 6800 GT/Ultra sported 16 pixel pipes), but the GeForce 8600 only has one fourth of the GeForce 8800’s resources when it comes to processing power and half of texture units.
The general architecture of the G84 is similar to that of the G80, but there are some differences too. Each stream processor (SP), or ALU, can perform two simultaneously issued scalar operations like MAD+MUL per cycle, just like a stream processor from the G80. In order to compensate the relatively low amount of stream processors, Nvidia pushed their clock-speed upwards to 1.45GHz on the GeForce 8600 GTS and there are signs that the company may be working on a more powerful solution with shader processor’s clock-speed maxed to even higher frequency. Besides, with the G84 the developer decided to change configuration of texturing units: Nvidia claims that the GeForce 8600 has 16 texture address (TA) and 16 texture filtering (TF) units, whereas the GeForce 8800 supports 8 texture address and 16 texture filtering units. It is unknown why the firm thought that rapid texture fetch would be important on the performance-mainstream and mainstream parts and why sacrificing “free” anisotropic filtering in favor of increased amount of TAs. We will have to come back to the claim about 16 texture units later in this review.
But while the GeForce 8600-series graphics processors do not feature as many execution units as the GeForce 8800-series, they still consist of 289 million of transistors, a respectable number. But the number has a logical explanation: the G84 graphics chip has built-in NV I/O logic, which alone consists of around 60 million of transistors and which needed to be taken out of the G80, as making a chip that consists of over 700 million of transistors using TSMC’s 90nm process technology seems to be too risky. Even though Nvidia’s original explanation about the reasons to move NV I/O logic off-die indicated that this was necessary to do because of extreme clock-speed of SPs, some may doubt now whether this was actually the reason: SPs on the G84 operate at 1.45GHz and output logic is still inside the GPU itself.
Since there are no major architectural difference between the G80 (GeForce 8800) and G84 (GeForce 8600), all the 3D capabilities available on the high-end part are present in the mainstream chip as well. Therefore, end-users can expect the following features:
- DirectX 10 and shader model 4.0 support
- High-quality angle-independent anisotropic filtering
- Coverage sample antialiasing
- HDR rendering with antialiasing
- Other features and peculiarities of the GeForce 8
But besides some tweaks within the pixel pipeline, the new GeForce 8600 also features enhancements not available on the G80: we are talking about substantially improved video processing engine.