PS2 vs. Xbox: GPU Technical Comparison
Posted: 2002-12-15 10:10am
RENDERING PIPELINES
A rendering pipeline is the largest part of the GPU; it is responsible for taking the raw polygon data and filling them out with textures. Consider these to be like the displacment of a car engine; the more pipelines, the more polygons can be pumped out in a clock cycle.
NV2A: This GPU is completely identical to a GeForce 3, but with two vertex shaders (we'll deal with the vertex shaders later). As such, it has four rendering pipelines. Each pipeline has two texturing units -- that means that for each clock cycle, each pipe can apply two textures to the same polygon. If two textures have to be applied to a polygon for special effects, this can be done without forcing the polygon to go through the pipe again. This is where Microsoft gets it's extremly high fillrate claim -- but having extra texturing units per pipeline has no impact on performance in the case of only one texture is needed for a polygon. One pipeline can put out a maximum of one polygon per clock cycle no matter how many texturing units it has.
Graphics Synthesizer: 16 rendering pipelines, but each one only has one texturing unit, so any special effects need to be passed through again. So, in the case that a polygon needs to have more than one texture applied, the effective throughput is cut in half.
LIGHTING AND TRANSFORMS
Lighting and transformation effects are what are responsible for many reflective surfaces, curved edges, and special lighting effects. This is important for performance, because doing it in software is very CPU intensive.
NV2A: Like the GeForce 3, it has full DirectX 8.1 compatability, and can support all lighting and transformation effects in hardware, freeing the CPU from having to do it. This is crucial, because the 733Mhz Celeron is much to slow to be able to handle this in software mode, and cause 3D performance to drop to almost nothing.
Graphics Synthesizer: No hardware support whatsoever; it relies on the Vector units to supply any special transformation or lighting effects. The Vecotrs are powerful enough for this kind of work without using the main CPU, but it is up to developers to write code specially to make this work.
PIXEL SHADING
The latest and coolest 3D manipulation effect introduced by DirectX 8 are the programmable pixel shaders. This allows developers to program exact shading and lighting effects for every pixel on a texture -- allowing for very high levels of detail.
NV2A: This is where this GPU differs from the GF3; instead of one vertex shader, it has two. Many effects of pixel shading require two passes of the vertex shaders, so having two will save the GPU from having to run a polygon through the shader a second time.
Graphics Synthesizer: Again, no native hardware support whatsoever. The Vector units are powerful enough to provide this, but it is up to the developers to code this in.
MEMORY BANDWIDTH
No matter how matter how many features or pipelines a GPU has, it must in the end rely on texture and geomotry data being fed to it on a regular basis in order to keep itself full and working at top speed.
NV2A: Features a 256-bit bus connected to the main system memory equalling 64-bits for each pipeline with a theoretical top performance of 6.4GB/s. It also supports hardware texture compression, which does cause a virtual increase in texture bandwidth.
Graphics Synthesizer: Has a massive 2,560-bit bus, but only 512-bits is assaigned for texture data to the pipelines for 32-bits per pipe. It is fed directly with a 4MB video buffer of 48GB/s, and beyond that by the main system memory of 3.2GB/s.
CONCLUSSION
The NV2A is the clear leader in features and native hardware support. All of the code nesicary to make use of these features is available to the developer.
The Graphics Synthesizer is a polygon rendering monster, capable or absolutely staggering power. Per every equalized clock cycle, it can output 16 polygons compared to the NV2A's 6.5 polygons.
A rendering pipeline is the largest part of the GPU; it is responsible for taking the raw polygon data and filling them out with textures. Consider these to be like the displacment of a car engine; the more pipelines, the more polygons can be pumped out in a clock cycle.
NV2A: This GPU is completely identical to a GeForce 3, but with two vertex shaders (we'll deal with the vertex shaders later). As such, it has four rendering pipelines. Each pipeline has two texturing units -- that means that for each clock cycle, each pipe can apply two textures to the same polygon. If two textures have to be applied to a polygon for special effects, this can be done without forcing the polygon to go through the pipe again. This is where Microsoft gets it's extremly high fillrate claim -- but having extra texturing units per pipeline has no impact on performance in the case of only one texture is needed for a polygon. One pipeline can put out a maximum of one polygon per clock cycle no matter how many texturing units it has.
Graphics Synthesizer: 16 rendering pipelines, but each one only has one texturing unit, so any special effects need to be passed through again. So, in the case that a polygon needs to have more than one texture applied, the effective throughput is cut in half.
LIGHTING AND TRANSFORMS
Lighting and transformation effects are what are responsible for many reflective surfaces, curved edges, and special lighting effects. This is important for performance, because doing it in software is very CPU intensive.
NV2A: Like the GeForce 3, it has full DirectX 8.1 compatability, and can support all lighting and transformation effects in hardware, freeing the CPU from having to do it. This is crucial, because the 733Mhz Celeron is much to slow to be able to handle this in software mode, and cause 3D performance to drop to almost nothing.
Graphics Synthesizer: No hardware support whatsoever; it relies on the Vector units to supply any special transformation or lighting effects. The Vecotrs are powerful enough for this kind of work without using the main CPU, but it is up to developers to write code specially to make this work.
PIXEL SHADING
The latest and coolest 3D manipulation effect introduced by DirectX 8 are the programmable pixel shaders. This allows developers to program exact shading and lighting effects for every pixel on a texture -- allowing for very high levels of detail.
NV2A: This is where this GPU differs from the GF3; instead of one vertex shader, it has two. Many effects of pixel shading require two passes of the vertex shaders, so having two will save the GPU from having to run a polygon through the shader a second time.
Graphics Synthesizer: Again, no native hardware support whatsoever. The Vector units are powerful enough to provide this, but it is up to the developers to code this in.
MEMORY BANDWIDTH
No matter how matter how many features or pipelines a GPU has, it must in the end rely on texture and geomotry data being fed to it on a regular basis in order to keep itself full and working at top speed.
NV2A: Features a 256-bit bus connected to the main system memory equalling 64-bits for each pipeline with a theoretical top performance of 6.4GB/s. It also supports hardware texture compression, which does cause a virtual increase in texture bandwidth.
Graphics Synthesizer: Has a massive 2,560-bit bus, but only 512-bits is assaigned for texture data to the pipelines for 32-bits per pipe. It is fed directly with a 4MB video buffer of 48GB/s, and beyond that by the main system memory of 3.2GB/s.
CONCLUSSION
The NV2A is the clear leader in features and native hardware support. All of the code nesicary to make use of these features is available to the developer.
The Graphics Synthesizer is a polygon rendering monster, capable or absolutely staggering power. Per every equalized clock cycle, it can output 16 polygons compared to the NV2A's 6.5 polygons.
