As of today two major commercial solutions control the market of 3D graphics specialized hardware: NVIDIA and AMD/ATI. ATI is behind the Radeon series while NVIDIA is responsible for the GeForce, Quadro and Tesla line. This chapter tries to cast some light on the state of the art of current available graphics hardware architecture perfectly aware, due to the rapid development of new solutions, how rapidly the proposed material can become obsolete.

3.1 - NVIDIA GT200 series

Figure 3.1: NVIDIA GT295 series (image courtesy of NVIDIA group).

The GeForce GTX 200 series represents the tenth generation of NVIDIA's graphics cards. Precursor to this series are the GeForce 8 and 9 series which share the same shader model architecture also known as Unified Shader Model (or Shader Model 4.0 under DirectX 10). Latest unveiled model of the GTX200 family are GTX285 and GTX295 released in January 2009. The 295 (see Figure 3.1) model introduces a dual GPU architecture in terms of a pair of updated GT200 (referred to as GTX200b) with 240 processors each boosting the total number of shader cores to 480. Each GTX200b chip features also 80 texture units, 30 render back ends (RBEs or ROPs) and 448-bit memory interface. It supports multi-display desktop mode, therefore enabling stereoscopic 3D visualization, increased multi-thread architecture and less power consumption. The 285 model is still a single GPU model represents a step up from the GTX 280. As the GTX295 model it relies on a GTX200b architecture which increases the performance by at least 15% with respect to its 280 ancestor.

Designed to support CUDA technology, the GTX200 architecture is claimed to be OpenCL compliant as well, with respect to future OpenCL releases.

3.2 - ATI Radeon R700 series

Figure 3.2: ATI Radeon R700 series (image courtesy of AMD/ATI group).

The ATI Radeon R700 series (see Figure 3.2) represents the graphics processing unit series released by AMD/ATI Graphics Product Group. Based on the RV770 chip its architecture extends the R600's unified shader architecture by increasing the stream processing unit from 320 to 800 units grouped into 10 SIMD cores. The RV770 chip features also 40 texture units and 16 render back ends (RBEs or ROPs) and 256-bit memory interface.

Designed for dual GPUs architectures and being the first chip designed to fully support GDDR5 memory the RV770 chip has been employed in the manufacturing of the Radeon HD 4870 released in August 2008. With a stream power of 640 shader processors the Radeon HD 4870 X2 features 2 GB GDDR5 memory, while the Radeon HD 4850 X2 still being a dual GPU architecture relies on a cut down version of the RV770 GPU referred to as RV770 LE adapted for a 256-bit GDDR3 memory interface.