   |
|||
|
|||
|
:
News
: Archives
|
|
NVIDIA’s GeForce 7800 GTX |
|
|
Walk Softly… |
|||
|
|
|||
|
by Josh Walrath |
|||
|
The G70 is comprised of 24 pixel pipelines, 8 vertex shaders, 16 ROPS, and all of the other architectural features that the NV4x series brought to the table (Pure Video, crossbar memory controller, compression techniques, etc.). At first glance the average consumer would assume that this is just an NV4x design with 8 extra pixel pipelines and 2 vertex shaders. This of course is not entirely true. NVIDIA took the SM 3.0 functionality of the NV4x and redesigned it at the pipeline level. The NV4x basic pipeline featured two ALUs and a texture unit attached to one of the ALUs. Theoretically each pipeline could do two shader ops or a shader op with a texture op. This of course was not always true, as one of the ALUs had a reduced instruction set. It appears as though NVIDIA has addressed this to a degree. Now it seems that both ALU’s have the full instruction set (or at least the inferior ALU now has a greater share of instructions), so they can theoretically do two complete shader ops at once. It does not appear though that the texture unit was decoupled from one of the ALU’s. Still, the G70 pipeline now has the ability to do two MADD (Multiply-Add) operations simultaneously, which the previous NV4x units were unable to do.
This is an actual photo of a wafer containing G70 chips. Even at 110 nm, each die is about 20% larger than the NV40 based GeForce 6800 GT/Ultra. MADD is one of the most commonly used shader operations in current applications, and doubling the theoretical MADD output should net some significant gains in shader heavy titles. MADD is not the only area that NVIDIA focused on. The overall pipeline design and its efficiency have gone through a major redesign. One example that NVIDIA gives is that a complex shader takes approximately 107 clocks to process on an NV4x pipeline. On the G70 that same shader takes 79 clocks to complete. This 35% gain in overall performance should go well with the increased clockspeed we will see with this product, not to mention the extra 8 pixel pipelines. Not every shader operation out there will see a doubling of performance, but it will undoubtedly perform faster on the G70 pipeline as compared to the NV40. The texture unit in each pipeline also received attention in the redesign. Texturing operations are now faster per clock as compared to the NV4x, and the cache design was also overhauled. The texture unit can now fetch texture data faster than its predecessor, and this has lead to improvements in anisotropic and HDR performance. NVIDIA has featured floating point texture units for some time, but this latest version is the fastest of them all. The vertex shader also received a lot of attention, which includes performance optimizations other than the addition of two more units over the previous NV40. The G70 features a faster triangle setup, as well as further optimized vertex units. This will of course enable a large performance boost in applications which are more vertex bound. Normal Map Compression is also included with the G70, and it will be compatible with ATI’s 3Dc compression technology. Also, the G70 will be the first WGF 1.0 (Windows Graphic Foundation- the initial graphical component of the upcoming Longhorn operating system from Microsoft) compliant graphics chip on the market.
Next: Improvements in Antialiasing If you have found this article interesting or a great help, please donate to this site.
Copyright 1999-2005 PenStar Systems, LLC. |
|||
