The 8th of November 2006 saw NVIDIA drop the proverbial bomb, unleashing G80 onto the world. An almost complete departure from every base architecture they've ever released, G80's clustered scalar array processing model was the basis for two 8-series GeForce launch SKUs: 8800 GTX and 8800 GTS.
8800 GTX featured 768MiB of GDDR3 memory and a full G80 clocked at 575MHz core, 1350MHz shader and 900MHz mem. Taking two 6-pin power inputs from your PSU, it's the fastest and best looking consumer graphics product yet seen. 8800 GTS shares the very same good looking pixels, but uses G80 in slightly disabled form (96 SPs from 128, 5 ROP partitions from 6) and with only 640MiB of GDDR3 memory. Clocks are 500/1200/800, using the GTX notation above and it needs only one power input.
Priced at $399 MSRP now (and available for less), 8800 GTS was the most affordable G80-based GeForce product on the 6th Nov last year. Today sees NVIDIA drop the barrier to D3D10-supporting entry even further with a minor evolution of the 8800 GTS, and a renaming of the launch 8800 GTS so that customers don't get confused. You see today's GeForce 8800 is still called the GTS, but it ships with half the memory. Let's explain what's going on.
Current 90nm G80 die production and demand for 8800 GTX and 8800 GTS being what it is, NVIDIA have decided that there's room in the market for a GTS example with just half the memory. The clocks, PCB, heatsink and abilities are the same, just the DRAMs are different and the framebuffer therefore halved. With a ROP partition disabled, meaning two less DRAMs than a GTX to connect, GeForce 8800 GTS 320MiB boards currently use 10 32MiB (256Mib) Hynix HY5RS573225A devices, and the board we have pictures of (overclocked XFX XXX Edition) uses the 1.1ns version at that.
Indeed, we expect most 8800 GTS 320MiB boards to come with 900MHz capable memory, so some overclocking from the standard 800MHz on reference-clocked boards should be doable for the adventurous.
Rated to the same 900MHz as the DRAMs on a GeForce 8800 GTS 640MiB (as it's now called), 8800 GTS 320MiB trades on the fact that at common resolutions that it'll be asked to drive, its memory pool is more big enough for rendering with good levels of IQ enabled.
The boards themselves will look no different to 640MiB variants unless you take the cooler off and peek at the DRAMs, the 8800 GTS 320MiB that we've looked at using those Hynix memories and a G80 core produced in the dying weeks of '06.
At the 1280x1024 and 1680x1050 (popular 17", 19" and 20" LCD panel) full-screen resolutions NVIDIA think a GeForce 8800 GTS 320MiB will be asked to drive more than anything else, the memory pool size shouldn't impact performance too much. While we've yet to round up the full gamut of results for our 320MiB review sample, we can say that that even up to 1920x1200 (popular 24" panel size) that the 320MiB GTS will be within spitting distance of the 640MiB variant at the same clocks in most (there are one or two outliers that like more than 320MiB of framebuffer) games with decent levels of AA and AF and that core and shader clock play more of a part in GTS performance than the framebuffer size, at least at this point in time.
Consider the case of a 1920x1200 fullscreen app doing 8xMSAA with a FP16 backbuffer, with one 2Kx2K shadow map, one 512x512x152; cube map, and a 32bpp Z-buffer. Including a guess at around ~30MiB of MIP chain data for the aforementioned maps, that's roughly 150MiB of memory cost for those items per frame, leaving you about the same again for geometry and texture data (and anything else) on a 320MiB board. Unless the app goes crazy with material detail, a memory pool that size should somewhat suffice for a modern game. Users of lesser resolution screens will suffer even less, and there's the usual scaling back on AA to help, too (arguing that the GTS 320MiB isn't a 2560x1600 product of course).
Add to that D3D10's virtualisation of the entire memory space in terms of where resources are created and how they're used on the hardware, graphics solutions with less memory get larger benefits because of it, given muchmore resource virtualisation in that API versus D3D9.
Therefore, at $299 MSRP, NVIDIA seem to have a very appealing sales winner on their hands. Assuming performance at common resolutions that's very much equivalent to the 640MiB version, the 320MiB sibling makes a good deal of financial sense for most.
We'll have a full look at GeForce 8800 GTS 320MiB in due course, and from a manufacturer most won't have heard of no less. In the meantime, the Hexians have taken a look at a couple of boards, from XFX and Foxconn.