NVIDIA's GeForce2 GTS GPU is clocked at 200MHz and, in this roundup, is matched to 32MB of 333MHz DDR SGRAM via a 128-bit memory bus. This provides about 5.3GBps of memory bandwidth. The GeForce2 GTS GPU can render four pixels per clock with up to two textures per pixel, which yields an 800Mpixels per second fill rate and a 1600Mpixels per second texel rate.

The GeForce2 GTS has a feature called the NVIDIA Shading Rasterizer, which is essentially a programmable pixel shader. A programmable pixel shader allows developers to perform programmable per pixel shading effects in hardware. The most impressive per-pixel effect we have seen is per-pixel bump mapping.

The GeForce2 GTS also includes a hardware transformation and lighting engine capable of theoretically rendering over 25 million polygons per second. Hardware T&L can help render highly complex scenes and removes the burden of T&L calculations from the CPU. Many current games gain some benefit from hardware T&L, but few games gain a massive benefit and none we know of require hardware T&L. For a more in-depth look at the technology please refer to our guide.

3dfx's Voodoo5 5500 AGP uses two 166MHz VSA-100 processors working in concert to draw a 3D scene. Each card carries 64MB of 166MHz SDRAM on a 128-bit memory bus, providing a total of 5.3GBps of memory bandwidth. Each VSA-100 can render up to two pixels per clock with one texture each, and with two VSA-100s, the Voodoo5 5500 AGP yields 667Mpixels per second fill and texel rates.

The VSA-100 has a feature called the T-Buffer (which you can read about in-depth here), which acts as an accumulation buffer. The T-Buffer gives developers the option to perform calculations on an image after it is sent through the graphics pipeline without having to send the data through again. T-Buffer effects include but are not limited to FSAA, depth of field, motion blur, and soft shadows.

The Voodoo5 5500 AGP does not include hardware transformation and lighting.

The RADEON GPU (which is disected in our technology guide) is clocked at 166MHz and is paired to 32MB of DDR memory also running at 166MHz (as opposed to 183/183MHz on the 64MB card). With two pixel pipelines and three texture units, the Radeon can process up to six texels (textured pixels) per pass giving the Radeon a 333Mpixel and 1Gtexel fillrate. The transformation and lighting engine was originally targeted to render 30 million polygons per second, but after a decrease in core frequency this number has dropped to approximately 25 million.

Like the GeForce2, the Radeon also makes use of programmable pixel shaders. While the GeForce2 supports functions with two textures and two operations per clock cycle, the Radeon supports three textures and three operations per cycle. ATI has also incorporated a proprietary feature called HyperZ to save memory bandwidth on necessary calls to the Z-buffer, raising effective fillrate. Four-matrix skinning and keyframe interpolation round off a feature-rich package making the Radeon our first DirectX 8-ready card.