GeForceFX Preview

Memory :

A lightening fast GPU is absolutely pointless without some lightening fast memory to partner it. DDR has served the industry well for some time now but it has had its day and it's time for something new. That something new is DDR2! Although we have been told that the GeForceFX will ship with 1000MHz DDR2 running at 2.0ns we'll no doubt see variants of the NV3X ship with 2.2ns 900MHz memory and presumably at some point with 1.8ns 1100MHz DDR2 also.

Contrary to what many people were speculating NVIDIA have decided to stay with a 128bit memory interface for the time being. Clearly they feel the bandwidth afforded them by DDRII along with the bandwidth saving features they've incorporated give them enough headroom for at least this current generation of products.

I'm assuming the memory actually used will be Samsung's K4N26323AE-GC20 (and I could equally be totally wrong) but if it is the specs are as follows.

Features

The figure quoted to us for memory bandwidth was 62GB/sec which is an impressive figure but we must remember this is effective bandwidth rather than an actual bandwidth. I'm currently waiting for confirmation as to the actual physical bandwidth. As with DDR we're looking at a conventional 144pin FBGA package so physically it's not going to look any different.

The memory architecture is being billed as "3rd generation Lightspeed Memory Architecture". This may well change as NVIDIA tells us it's significantly different to previous incarnations of their Lightspeed memory architecture but no precise details on these differences were forthcoming during our conference call or evident in the technical papers. The GeforceFX is currently the world's only DDRII production GPU and its >1GHz data rate is also a world first.

 

DirectX 9 and GeForceFX
Shader Functions

shader programmability: The DX8 pixel shader specifications could appropriately be termed “configurable” rather than programmable because the DX8 spec needed the general programmability associated with a full instruction set and flexible programming structure. The GeForce FX GPUs and DX9 allow much longer shader programs, and give developers a greatly expanded pixel shading instruction set.
Vertex shader programmability: The DX8 vertex shader specification gave developers very little control over program flow. DX8 vertex shader programs are executed linearly, with no early termination for performance optimization. DX9 and the GeForce FX GPUs support conditional branching for greatly improved program flow control.
Unified pixel shading specifications: The addition of Pixel Shader 1.4 to the DX8 specification created an alternate pixel shader structure and programming structure that followed a completely different implementation philosophy. Many developers were forced to re-write shader programs for different hardware, rather than writing strictly to the API. The DX9 specification eliminates this problem with the new unified Pixel Shader 2.0 definition.

The combination of DX9 and GeForce FX GPUs narrows the gap between state-of- the-art PC graphics and state-of-the-art rendered movies such as the Disney/Pixar Monsters, Inc. and Columbia Pictures Final Fantasy: The Spirit Within productions. The DX9 API is a critical enabling technology for this powerful new combination of hardware and software.

The DX9 specification includes three major new features:

Pixel Shader 2.0: DX9 exposes true programmability of the pixel shading engine. This makes procedural shading on a GPU possible for the first time.
Vertex Shader 2.0: DX9 dramatically enhances the power of the previous DirectX vertex shader by increasing the length and flexibility of vertex programs.
High-precision, floating-point color: DX9 breaks the mathematical precision barrier that has limited PC graphics in the past. Precision, and therefore visual quality, is increased with 128-bit floating-point color per pixel.

 

GeForceFX Pixel Shader 2.0+

As you might have guessed from the "+" in the name Nvidia's implementation of the GeForceFX's pixel shader goes beyond what's strictly supported by DirectX 9. Provided programmers are prepared to take advantage of what's on offer NVIDIA have really opened the floodgates on some potentially awesome effects. With an incredible 1024 texture and 1024 colour instructions available in addition to 16 textures per pixel NVIDIA have also massively increased the limits for shader instructions, constants and registers (see below). Forget about things like swizzling and conditional write masks, these are functions directly linked to the source registers and aren't really something you need to be concerned with unless you're actually programming the shader.

Scientific Instructions :

Before DirectX 9 and GeForceFX, pixel shader functions were limited to a very restricted array of functions and were considered to be configurable rather than truly programmable. Pixel shader 1.4 specifications of DirectX 8 were primarily limited to texture related operations such as fetching, blending and so on. Pixel shader 2.0 introduces several new functions to the fray including maths instructions that allow it to solve arbitrary math algorithms. One example is the alien below where the plain blue version on the left can be transformed to a stylized pink/brown version simply by making the colour and shade a function of the angle at which the light hits it. With a little imagination from the programmer simple maths procedures can be made to create some fairly impressive effects.

Flow Control :

Another key element to the pixel shader 2.0 armoury is the introduction of more efficient and thus more productive flow control handling. For the first time in Direct3D, DirectX 9 introduces functions such as branching and looping to the pixel shader specifications. This allows code to be routed and reused like never before and also simplifies the creation of effects as they can be grouped into categories and called upon whenever they're required. This drastically cuts down on development time as large quantities of the code can be shared between numerous effects and can be debugged in a more efficient and generalized manner. Programmers can now develop a a generalized wood-grain shader, a generalized vegetation shader, a generalized eyeball shader, etc.

Procedural Shading :

The improved flexibility of the pixel shader 2.0 arbitrary maths functions greatly enhances what can be achieved. If we look at procedural textures for an example programmers can now develop more complex 2D and 3D patterns on demand without the need for a texture map. If we liken this to a cake, the old method involved trudging down to the shops, picking up your cake and trudging home with it again. This is obviously labour intensive and nor particularly efficient. The procedural approach is that rather than going to the shop for the cake you'd simply slip a recipe into your oven and the cake would be created based on the instruction contained in the recipe. This approach saves on resources all round and totally streamlines the whole procedure. What Pixel shader 2.0 does is allows far more lines to be included in the cake recipe and thus more elaborate cakes can be conjured up.

If we look at the image below, the amazing thing is that no texture maps at all were used in its creation. Every surface from the wood-grained table top to the rough surface finish on the vase were computed mathematically and no traditional texture maps were required at all. Another benefit to this technique comes in games where perhaps a gunshot would blast away the surface of a textured object like a brick wall or marble pillar. One way of achieving this is by using 3D textures which not only define the external appearance but also texture the internal area of the object. This method is generally not very efficient as large 3D textures are being shippped over the internal bus despite the fact they may never actually be seen, depending how good a shot you are of course! With procedural textures the surface that remains after the surface has been blasted away would simply have a new texture created for the underlying surface mathematically, a far more efficient way of handling it.

 

GeForceFX Vertex Shader 2.0+

The GeForceFX vertex shader enhancements closely mirror those of the pixel shader. Vertex Shader 2.0 introduces true programmability to the DX9 vertex shader specification. Two key changes, compared to the DX8 specification, are longer vertex programs and flow control using conditional branching and subroutine commands. Vertex Shader 2.0 has also extended the instruction set with new math functions.

With this new flexibility programmers are free to write more complex routines for sophisticated and realistic effects with less system and development time overhead.

DX9 vertex programs also simplify the animation of
complex geometries such as water

While the vertex shader is calculating the deformations required to animate the water's surface the pixel shader is busy calculating things like refraction on a per pixel basis. By accurately showing the way light either shines through or is reflected off the water's surface the sense of realism is greatly enhanced.

Matrix Palette Skinning :

An example of how the GeForceFX's advanced vertex shader 2.0+ can let's take Matrix Palette Skinning as an example.

DX8 / NV2x / R200 / RV250 4 shaders 1 bone 2 bone 3 bone 4 bone Segment Model into those polys depending on 1,2,3,4 bones Draw separately

DX9 / Radeon 9700 1 shader (1-4 bones) Branching is per-object Still have to segment model into 1-4 bone groups Draw separately

GeForce FX 1 shader - branching is per-vertex No need to segment model Loop is done conditionally on a pervertex level

Because the GeForceFX is able to perform branching on a per-vertex level rather than per-object a four segment skinning operation can be performed in a single operation with no need to draw each of the bones independently.

In case you're wondering who the good looking character above is, his name is Ogre and he features in another of NVIDIA's demos. This "Dancing Ogre" is a real-time rendition of a movie originally created by Spellcraft Studios titled "Yeah! the movie". The fidelity of the characters in the scene (polygon counts, materials, and complexity of motion) are nearly indistinguishable from that of the original off-line rendered composition.
Check out Spellcraft Studios from whom the original material to produce this demo was provided.

 

You may also remember our friend The Wolfman from the GeForce4 launch. Despite being a major step forward the GeForce4 combined with DirectX 8 still had to render the fur as individual layers of geometry using eight passes for each individual pixel. Thanks to DirectX 9 the GeForceFX is now able to perform the same operations in a single pass leading to huge performance gains and no subsequent loss in quality!

 

Time Machine is one of the demos NVIDIA will be using to demonstrate the power of their pixel shaders. Here's NVIDIA's explanation of what we'll see :

Gaze through a portal at the passage of time as this 1950's era pickup truck shows the ill effects of decades of neglect. Moving from its pristine condition in 1950 to an old rust bucket of today, the power of the programmable GeForce FX pixel engine blends a variety of material surface effects into a single shader program. The combination of procedural and high-resolution texture data creates a series of realistic surface treatments on the paint, chrome, wood, and interior components of this old workhorse.

Key Feature:

Time-based Shaders - Each of the aging materials has a single pixel shader associated
with it. These shaders use a variety of texture map inputs (color, bump, specular,
reflection, surface reflectivity, and reveal maps) to produce a seamless transition of
surface material effects over time.

 

<<< Last Page | Summarizing So Far>>>

Select Destination Page Introduction, AGP8X Interface Colour Precision Colour Precision Cont. Memory, Pixel/Vertex Shaders Summarizing The Features FSAA, Cg, UDA and Board Features

Home