GPU NVIDIA® GEFORCE™ FX 5700 Ultra

• Bus Type AGP
• Memory 128MB
• Core Clock 475MHz
• Memory Clock 900MHz (effective)
• RAMDAC Dual 400MHz
• API Support Microsoft® DirectX® 9.0, OpenGL for Microsoft® Windows®
• Connectors VGA, DVI, S-Video out
• 356 million vertices/sec.
• 14.4GB/sec. memory bandwidth

The 5700 Ultra is based upon NVIDIA's NV36 chip, which is based upon their GEFORCE FX 5950, not their GEFORCE FX 5800, which their last mid-range card was based upon. The 5700 Ultra is based upon completely new architecture from their last mainstream chip, the 5600 Ultra which is based upon the chip codenamed NV31, the NV36 brings new features to NVIDIA's mainstream video card chip sets which are reflected here.

NVIDIA implemented what they call CineFX 2.0 with the GEFORCE FX 5700 Ultra. This architecture features support for Pixel Shader 2.0 instructions (we'll look at PS performance later in the review). In fact, in some ways the CineFX 2.0 architecture goes beyond the specification of Pixel Shader 2.0. The GEFORCE FX 5700 Ultra supports up to 1024 instructions in a single pass in the pixel shader. The architecture also supports loops branches and other features which aren't of much use right now. But what's important is performance, not theoretical features.

First is the much improved vertex shader pipelines over the previous generation. The GEFORCE FX 5600 had only one vertex shader, which limited its usefulness with high polygonal games. The 5700 Ultra, on the other hand, has three parallel vertex shaders giving a vertex thoroughput of 356 million vertices a second. This is over 3x the thoroughput of the 5600 series of cards, and equals the vertex thoroughput of its bigger brother, the 5950 Ultra.

Fillrate is important in today's video cards as well. The 5700 Ultra is a 4 pixel pipeline card. With a core clock speed of 475 MHz, this gives a theoretical maximum fillrate of 1.9 Gigapixels a second (500 MHzx4 pixels a clock). It's going to be interesting comparing the 5700 Ultra versus the 9600 XT, as this card theoretically has less fillrate 2 gigapixels for the 9600XT versus 1.9 for the 5700 Ultra.

Another important factor is memory bandwidth. The 5700 Ultra has a memory clockspeed of 450 MHz DDR (note, that I never double memory clock speed, it's a personal preference). The memory bus is 128-bit which gives a memory bandwidth of 14.4 GB/second. The formula for memory bandwidth in this instance is: 450x2x128/8=14.4 GB memory bandwidth. This compares favorably with the 5600 Ultra of the last generation's bandwidth of 12.8 GB/second and the competition's (9600 XT) memory bandwidth of 9.6 GB/second. As to how it'll play out in real life, we'll have to wait and see.

VPU ATI Radeon 9600XT

• Bus Type AGP
• Memory 128 MB
• Core Clock 500MHz
• Memory Clock 300MHz (600 MHz effective)
• RAMDAC Dual 400 MHz
• API Support Microsoft® DirectX® 9.0, OpenGL for Microsoft® Windows®
• Connectors VGA, DVI, S-Video out
• 250 Million vertices/second
• 9.6GB/sec memory bandwidth
• Up to 16 textures per pixel per pass

The Asus 9600XT is based upon ATI’s RV360 chip, which is a refresh of their RV350 chip that formed the basis of the Radeon 9600 Pro. The 9600XT chip is a 4 pixel pipeline card, like the 5700 Ultra. The card also has 2 vertex shader pipelines compared to the 3 found on the 5700 Ultra. In terms of performance and real life gaming the two cards should perform similarly.

The 9600XT also has the feature Dynamic Overclocking, which is also found on the Radeon 9800XT announced a week earlier. With this feature, a thermal diode reads the temperature of the 9600XT. If the temperature reads higher than a certain degree, the card is clocked at a lower core clock speed. If the temperature is lower than that degree, the higher clock speed is used. It’s fully covered by the warranty. If you don’t feel like using Rage3d Tweak or Powerstrip, it’s a free overclock of the video card.

ATI terms their pixel shader and vertex shader architecture SmartShader. The version of Smartshader supported on the 9600XT is 2.1. The card is capable of 160 pixel shader instructions in a single pass, 32 texture address operations 64 vector operations and 64 scalar color operations in a single pass. Microsoft’s DirectX 9.0 specification requires 96 operations in a single pass. Of course with multi-passing more instructions can be used. The Vertex Shader of the 9600XT is capable of 65,000 operations in a single pass with loops, and subroutines.