Historically weâ€™ve been used to high end graphics refreshes every 6-9 months, however its been some 14 months since NVIDIA initially announced their Shader Model 3.0 NV40 architecture and there has been no refresh of the product since then. Of course, the need for refreshing were somewhat sidelined with an architecture that has the latest shader capabilities and SLI providing a mechanism for improving performance should end users wish to. Its not as though any of the NV40 based GeForce 6800 products were fairing poorly, performance wise, over the past year of availability either.
None the less, it is always an inevitability that new products will eventually become available and NVIDIA have tried to keep people guessing a little by the change of internal codenames. Initially indications were that NV47 would be NVIDIA's next high end chip update, with suggestions of it increasing the number of internal pipelines and being based on a 110nm process. Of course, before "NV47" could come to the market codenames of G70 and G80 surfaced which possibly threw people off the scent.
Now, NVIDIA is ready to unveil and release G70 (G = GeForce, 70 = 7 series) in its initial form of GeForce 7800 GTX. Here we'll take a look at the architectural changes for G70 as well as taking a look at the 7800 GTX board and it performance.
With the change from the NV30 to NV40 architectures NVIDIA altered their architectural philosophy to one that ATI had already proven fairly successful with their Radeon 9700 products. NV30 featured a somewhat deep, 4 pipeline architecture, with multiple ALU's (Arithmetic Logic Units) in the fragment (pixel) shader pipeline and the register combiner in there as well; NV40 cut the depth of the pipeline down to have two ALU's, also tasking one with some texture duties - the upshot of trading off pipeline depth was to dramatically increase parallelism by moving from 4 pixels pipelines (a single quad) to 16 pipelines (4 quads of operation). G70 follows in the same path as NV40, but takes the parallelism a further step on.
As the diagram above indicates, the fragment shader pipelines have increased to 24 (6 quads of operation), however the ROP's remain at 16, meaning that the maximum pixel sampling per cycle will be 16 colour or 32 Z/Stencil operations. This type of fragment shader and ROP arrangement is very similar to the arrangement of NV43 (GeForce 6600), although in this case there are obviously a lot more pipelines. As we highlighted in our GeForce 6600 GT article having a 1:1 fragment shader to ROP arrangement is becoming increasingly less important as the onus is being placed more on shader processing so the graphics workload is spending more cycles operating internally, within the shader pipelines, and actually sampling to the frame-buffer is not necessarily the most frequent operation now. Also, as G70 sticks with a 256-bit memory bus there would likely be limited gains should the ROP count be increased further anyway, hence it can be more useful to spend that transistor budget elsewhere.
Along with increasing the fragment shader pipeline count, G70 also increases the number of vertex shader units to 8.