updated 11:20 am EST, Mon January 18, 2010
NVIDIA Fermi is ultra-general, 68pc faster
NVIDIA edged closed to the formal release of boards based on its Fermi architecture with fuller details about the processor as well as an early test. The 512-core design is now known to also be a high-end graphics design and adds 16 geometry and 4 raster units that weren't present in previous GeForce 200 cards. These newly dedicated units let a Fermi chip better handle tasks out of order.
At the high end, Fermi will also use GDDR5 memory and has twice the available bandwidth even with a narrower 384-bit bus versus the earlier 512. Image quality should be helped by the presence of jittered sampling, which provides softer shadows, and both the Coverage Sample and Transparency Multi-Sample antialiasing have been rendered more accurate.
3D Vision Surround will give NVIDIA cards their first multi-monitor continguous display technique and should let up to three screens merge together to form one image.
Although the general-purpose computing is in many ways similar to the underlying GeForce 200 line, NVIDIA expects about an eightfold boost to CUDA, DirectCompute and OpenCL through several improvements. In addition to including its own Level 1 and 2 caches, Fermi can switch between graphics and general computing 10 times faster than before.
While full performance benchmarks haven't yet been revealed, the company has been willing to show Far Cry 2 running at high resolution at 84FPS, or about 68 percent faster than a top-end GTX 285. At CES, NVIDIA showed Electronista a demo with that used tessellation (a feature specific to DirectX 11 and newer OpenGL builds) to show an extreme level of detail in stone paths and other areas that would normally have to be simplified.
NVIDIA still hasn't outlined the cards that will use Fermi, but these should be announced, if not shipping, before the end of March. Initial boards are likely to keep the GeForce branding. [via AnandTech]