After a solid few weeks of gaming and benchmarking NVIDIA G70-powered GeForce 7800 GTX cards, I’m here to tell you what I think. First, let me get the obvious commentary out of the way: These puppies are wickedly fast. So much faster and more efficient than the previous 6800 series of GPUs from NVIDIA, that under the right conditions, a single 7800 GTX can outperform a pair of 6800 Ultras in SLI mode. By that I mean don’t spend $1,000 on graphics cards for previous-generation DX7 and/or DX8 class games. The 7800 GTX is geared for DX9 and then some at high resolutions with all of your favorite AA/AF eye candy turned on. It also eats more complex pixel shaders and is an AA monster. Clearly it’s built for upcoming DX9 games based on PS3.0 and FP32. The early demos I saw included Huxley, UT- 2007, and Age Of Empires III and looked absolutely stunning running with SLI. NVIDIA says that the 7800 GTX is not a “refresh” of the 6800, and instead the pixel and vertex shader pipelines are redesigned and gentrified, whilst at the same time decreasing power consumption. The core is still based upon a 0.11-micron process, and even though the clock speed is bumped up 5MHz to 430MHz, it has 302 million transistors (up from 222 million), indicating some serious intentions under the hood. The pixel pipelines are also up from 16 to 24, and even the vertex pipelines are up to eight. That’s some serious horsepower and torque, assuming that they were the same as the 6800s pipes, which of course they aren’t. Peak pixel fill rate remains at 6.8Gpps (gigapixels per second), but the pixel texture fill rate is now capable of 10.32Gtps (gigatexels per second), which is up from 6.8Gtps on the 6800. Using a 256-bit memory interface, the clock speed for the GDDR3 memory is up to 1.2GHz, and peak memory bandwidth is up from 35.2GBps to 38.4GBps, which takes over from the previous champ (Radeon X850 XT PE) that boasted 37.8GBps. But all those numbers that don’t sound too impressive, save the fact that efficiency in the graphics pipeline is improved, NVIDIA actually spec’d out over 1,300 different shader operations from games past, present, and future to find out which were the most popular, required optimizations and could be improved by redesigning the shader units. Obviously NVIDIA isn’t overly free with its secret sauce, but some things have been revealed, such as the scalar unit in the vertex pipeline that’s now 20% more efficient and the vertex texture fetch unit that’s received a boost. Also, the floatingpoint- texture processors are now jazzed up, which improves HDR capabilities.
Also, the FP32 shader units are now able to cope with two Vect4 MAD (multiply-add) operations per clock. Visually, there’s a new AA mode, called Transparency Adaptive AA, which takes AA further on to alpha blended areas that previously didn’t get the AA goodness in more traditional AA modes, such as multisampling. So, for example, all those chainlink fences set one behind the other that used to look blurry in Half-Life 2, now look more realistic. One area in which I was particularly excited to see NVIDIA move forward is video acceleration features, dubbed Pure- Video. ATI has long been king in that department, but now NVIDIA has listened to end-user and OEM feedback and stepped up to the plate. The programmable video processor and its HDTV and MPEG-2 features are accelerated and sport 2:2 pull-down correction. The only issue is that I’m still awaiting the Force- Ware Release 75 driver in order to dabble, test, and play with PureVideo. And I’m especially happy to hear about its apparent CPU-utilization tweaks. I look forward to seeing more games with DX9c-style Shaders, normal maps, HDR rendering, and floating-point textures. Clearly the best solution available is the 7800 GTX. ATI delayed its next-gen R520 architecture, but even so, it’s going to take something mighty special from the company to even catch up to NVIDIA.
