: News
: Reviews
: Editorials

: Archives
: About
: Contact
: Advertising
: Privacy
: Links
: Forums




NVIDIA GeForce 8800 GT


Revitalizing the Midrange


by Josh Walrath


            For the past two years the midrange market in the video card industry has had some ups and downs.  Products like the Radeon X1950 Pro helped redefine the area, but with the advent of DX10 and the introduction of the 8600 GTS and HD 2600 XT, the upper midrange has suffered somewhat.  Products such as the 8800 GTS 320 did help to add a bit of excitement to the scene, but problems with how it was able to address the 320 MB of memory caused it to be unpopular in some circles.

            This holiday season is something different though.  Last month NVIDIA released the 8800 GT, claiming that cards will be available from $199 to $249.  Then AMD released their HD 3850 at $179 and the HD 3870 for $219.  Suddenly, the upper midrange market was redefined with excellent products from both sides.  Of course, things have not quite worked out as many have hoped so far, but the fact of the matter is that we have some very interesting cards out at the moment.


The 8800 GT

            NVIDIA launched the first volley in what is going to be a highly contested Holiday Season.  The 8800 GT was not exactly a surprise, but its performance was unexpected considering the price points they were aiming at.  Original pricing was estimated to be $199 for the bog-standard 8800 GT with base clocks, and $249 for the overclocked and specialty editions.  So far that has not exactly panned out for consumers, but it is a good point to aim for.

            Before heading into the specifics of the card, I thought I would talk a bit about the G92 chip that this card is based on.  The G92 is fabricated by both TSMC and UMC on their respective 65 nm processes.  The 8800 GT features a G92 chip that has portions of it disabled.  The G92 chip features 128 stream processors, 64 texture address units, 64 texture filtering units, and 16 ROPS/Multi-Sample units.  In the 8800 GT one shader cluster with its associated texturing cluster is disabled.  So the 8800 GT features 112 stream proessors, 56 texture addressing and filtering units, but keeps the 16 ROPS.

            Previously I believed that the G92 chip had extra ROP clusters disabled, but that does not appear to be the case.  If it would have had more ROP clusters, it would have also had support for a wider memory pathway.  As it is, each ROP cluster has a 64 bit path to memory, giving it a total of 256 bits.  Each chip has less potential memory bandwidth than the older G80 chip, and it also has less raw pixel fillrate and multi-sampling ability than the older chips.

            The G92 chip is comprised of 754 million transistors filling up around 324 mm square of die space.  Compare this to the older G80 with around 680 million transistors taking up 484 mm square of die space, and we can see that NVIDIA is really packing in a lot of trannys in a small amount of space.  No San Francisco jokes please.  While cutting down the ROP hardware (enough tranny references) NVIDIA added a bit more functionality to the chip.  The biggest change is updating the PureVideo unit to PureVideo 2.  PV2 adds another level of CPU offload when doing VP1 and H.264 decoding.  PV2 does not do as much offload as AMD’s UVD (Universal Video Decoder), but it does cut down the CPU cycles quite a bit nonetheless.

            Texturing power is also enhanced over the older G80.  The G92 follows the G84/86 parts by doubling the amount of Texture Address units in each cluster.  This gives the G92 a huge amount of texturing throughput.  Texture Filters remain the same as the G80, so overall filtering ability is improved only through the higher core clockspeed of the chip.  NVIDIA also integrated the NVIO-1 chip, which was previously separate from the G80.  This handles the output capabilities of the G92 chip, and it supports HDCP at 2560 x 1600.

            The final piece of joy that was added to the chip is the ability to communicate on the PCI-E Gen 2 bus.  This doubles the bandwidth over PCI-E Gen 1, and theoretically will improve performance on applications which require a lot of bandwidth between the GPU, CPU, and main memory.  So far not many apps have maxed out even an 8X PCI-E Gen 1 connection.  SLI throughput should get a decent boost though in applications which will stress that, but according to NVIDIA most of the SLI data is actually transferred through the “over the top” connection.  Still, it probably can’t hurt.


Next:  More G92 Details


If you have found this article interesting or a great help, please donate to this site.


Copyright 1999-2007 PenStar Systems, LLC.