Thoughts
A quick note about Doom3. While I didn't explicitly show results in this article, 6600 GT's performance is usefully above what the 5900 PCX could manage. Its basic architecture tells you that's going to be the case without even testing it, given NV40's baseline performance. So if you're looking at the 6600 GT to help you play a nice game of Doom3, go ahead and grab one, it does well, even at my LCD's native resolution of 1600x1200. I have to drop visual quality due to the 128MB of memory on the board but it's doable while still looking good.Besides Doom3, performance was exactly where it should be, given the way NV43 is setup. In older DX8-class tests when we're more likely to be texture fillrate limited, it's not that much faster than a 5900 PCX at XT clocks. NV35 has two texture samplers per pixel pipeline, giving it a fine texel fillrate figure for its clock speed. That's especially true at high resolution where the memory bus width and memory density (remember, most if not all 6600 GTs will ship with 128MB of memory) conspire to limit performance.
However, when we're shader limited (mostly pixelshader), it comes to life. We get our 60% delta compared to 6800 Ultra and given its eight basic pixel pipe setup and three vertex shader units (half NV40 in those basics), paired with a high core clock, that was roughly what you could have predicted before sitting down and doing any testing at all.
NVIDIA have designed NV43 well. They realised that in titles that really stress NV4x (titles that NV3x struggle with), 128-bit memory bus and a cut down ROP count are great ways to save die space and transistor count without impacting performance too much. It's a balanced GPU, setup so that it can operate well despite only being able to write four pixels per clock to the output buffer, after processing up to eight pixels per clock internally.
That 60% delta is important when considering SLI. 6600 GT is the first mainstream GPU from NVIDIA to feature SLI capabilities. For those of you that've been hiding under a rock for the last few weeks, NVIDIA's SLI multi-GPU technology allows you to install up to two identical PCI Express graphics cards in a compatible system, linking them together for a large hike in performance. With nearly 90% extra performance on tap, according to NVIDIA, that 60% difference in raw performance between a 6600 GT and the current pinnacle of single-card NV4x performance, 6800 Ultra, means that a pair of SLI 6600 GTs should eclipse the performance of a single 6800 Ultra at 400MHz. And there lies the appeal. You drop in one 6600 GT now and another down the line for less money, getting NV40 Ultra performance for less money than you'd have spent on that bare board. Remember the target price of 6600 GT is some 60% of a 6800 Ultra at the time of writing. Good performance now, great performance later, less expense.
How that pans in out in real-world use, with market prices, SLI board availability and the actual performance benefit from SLI remains to be seen, but the theoretical performance is there. Enough to hopefully eclipse a basic 6800 Ultra at least. Any less and SLI loses some base appeal for the upgrader playing in the mid-range.
An AGP version (that can't do SLI however, for obvious reasons) and general retail availability across the board is all we wait for. Enthusiasts appear to be gagging for the general sale of 6600 GT and as we've found out today, for generally good reason. The quicker it comes to market the better, outside of OEM.
X700 Pro and the differing X700 XT (eight pipe R420 by all accounts) are the obvious contenders and will be available soon, so a performance match-up with those Radeon variants will be very interesting.
Very strong mid-range performance wth many times the raw shader power of NVIDIA's outgoing mid-range hardware, keen price and all the architectural benefits of NV40, without any of the form factor downsides of NV40 boards. 6600 GT will sell very well and for good reason. Get your pre-order in now and pray NVIDIA get the AGP version out as soon as possible since that's the version many of you are waiting for. It's the board to have on PCI Express at the time of writing.
Now there's only one piece of the pie left.
