It seems that ATI is well ahead of NVIDIA when it comes to GPGPUs. Still, NVIDIA is working hard on its CUDA design, which may be released with the new GeForce 8900 series.
CUDA actually stands for Compute Unified Device Architecture. It is similar to ATI's Firestream GPGPU project, apparently being a fundamentally new computing micro-architecture for the GPU, in order to solve complex computational problems across consumer, business, and technical industries. CUDA can't really work without an upcoming dual Geforce 8950 GX2 configuration, as these cards can deliver more computation power than current high-end server CPUs.
Although Nvidia entered the GPGPU market later than ATI, it looks like the Americans are quite happy with their G80 series general processing capabilities. As of now, G80 is still the fastest graphics chip on the market. It has 128 unified shaders and can process a lot of pixels. But those shaders aren't necessarily proficient at rendering complex graphics, as science calculations can easily be run on a graphics card disguised as a small shader program. Now, Geforce 8900 GTX is expected to have 25 percent more Shader power, and thus, dual Geforce 8950 GX2 configurations should provide a lot more Shader and Pixel power.
Chinese sources claim that the upcoming GeForce 8950GX2 card will sell for $599 USD. It is rumored to be the successor of the previous 7950GX2 and will come with a total of 512MB GDDR-4 memory per GPU. As part of the new 8900 series, NVIDIA plans to introduce the GeForce 8900GTX which will replace the current high-end 8800GTX. It will ship at a retail price of $549 USD and it will also see the inclusion of GDDR4 memory which allows NVIDIA to raise the clock speeds to outrageous 2200MHz frequencies. The power boost may also come from the new 80nm manufacturing process.
All this points to the fact that NVIDIA is already prepared for the AMD/ATI offensive. AMD/ATI is expected to take the covers off of their range of upcoming R600 graphics cards in a couple of weeks.