NVIDIA releases GeForce 8800GTX, 8800GTS, 680i, 680a

Shaders: A Final Unity
We mentioned the unified shader. NVIDIA has an architecture of 'steam processors' (SPs) – essentially what it's now calling the shader units. Whether Geometry, Pixel or Vertex shaders are passed on to the units is managed by, oddly enough, a thread manager named GigaThread. The earlier rumours that NVIDIA's pipeline wouldn't be fully unified are complete bunk.

The stream processors are scalar, meaning one pixel element can be handled per processor. So it would take four SPs to do the Red, Green, Blue and Alpha elements of a pixel. NVIDIA are claiming high efficiency of these processors, whether instructions are issued in vector or scalar, which is part of the reason for G80's new found speed. We're still waiting on details as to exactly how many instructions per clock cycle it can handle.

To compare, the pixel shaders in the Radeon X1K series take an input of 3 component vector (RGB) and one scalar (Alpha), and can process two vector and scalar instructions per clock cycle.

A traditional pipeline means things are treated sequentially as far as shaders are concerned – Vertex shaders are handled first, and then followed by pixel shaders. In a GeForce 7900GTX for example, there are 8 vertex and 24 pixel units. This is all well and good, but essentially means if a scene is vertex shader flooded then the pixel unit is only passed 8 vertices at a time to work on while the vertex unit tries to catch up, leaving 16 pixel processors doing nothing at all. The sequential nature also means that if a scene is pixel shader heavy it's still waiting for the vertex processor to feed it information.

Unified shaders are almost like the shared cache between CPU cores – a big block of SPs there for general use, and GigaThread dynamically allocates pixel, vertex, geometry, physics or GPGPU code where needed, meaning that it shouldn't matter whether a scene is vertex, pixel or geometry heavy. Whereas the total number of units the 7900GTX had were 32 combined, ATI had 56 on the X1900XTX (48 pixel, 8 vertex) the 8800GTX offers a maddening 128 units, double what ATI is rumoured to be offering on the R600.

Whether or not the stream processors can even be compared to ATI's unified shader processors remains to be seen. ATI may be running at a stupidly fast speed, or quite simply and more likely will have a completely different architecture. ATI has already tackled the unified shader front in the Xbox 360 and so have some experience, whereas G80 is NVIDIA's first shot. Then there's been talk of NVIDIA's part being a hybrid system, rather than fully unified – we wish we could confirm either way, yet NVIDIA stays silent on the topic whenever we ask them.

There's a new shader that's come to play in DX10 – the geometry shader. Whereas a vertex shader allows you to move individual vertices, and a pixel shader works on a per pixel basis, the geometry shader allows the GPU to alter an entire triangle and its surrounding vertices. This allows for some fancy and computationally fast deformations of meshes, and that surface can be tessellated – or basically, more triangles within that triangle can be created on the fly, feasibly to allow more realistic deformations with more gradual transitions, although no doubt other applications exist.

Geometry shaders aren't limited to distortion though, they can also be used for creation (or indeed, deletion). For example, a shadow shape can be extruded from a mesh, distorted appropriately based on light sources and shifted to be geometrically accurate. You could even take a vertex and generate entire triangles around it if you wanted to. The whole idea is simply an attempt to reduce the amount of data transferred between CPU and GPU – so an entire mesh can now be created from a single vertex on GPU, rather than the CPU attempting to jam the completed information down a limited bus.

G80 on a wafer. Mmmmm, wafer.

Other Bits
'Quantum Effects' is also here – it's NVIDIA's name for GPGPU physics. It appears to be more of a brand to say 'we can do physics on GPU' rather than being a discrete solution, although we're guessing that the stream processors are optimised to handle physics better than previous generations. Two titles, Hellgate London and World in Conflict were listed at the presentation as being upcoming physics enabled games – although later information has the first using HavokFX and DX10 physics, and the second using CPU – so we're unsure how this directly relates to Quantum Effects.

Also new is what NVIDIA calls 'Lumenex', essentially a watchdog for image quality. What this means to the average man is 128bit HDR, 16xAA on one card and finally HDR+AA capable hardware from NVIDIA. As part of this, NVIDIA also claims to have solved the 'shimmering' problem that affected complex textures such as a texture of millions of pebbles in its previous generation cards.

There is also a new AA mode available – application enhanced. Rather than the game or the driver deciding whether or not or what level of AA is applied, with application enhanced the application determines the AA status and what surfaces it is applied to, whereas the driver determines what level of AA is applied. This should help compatibility while giving access to greater levels of AA than the application has specified. The card is also set to use the newer CSAA (Coverage Sampling Anti Aliasing) over MSAA (Multi Sample Anti Aliasing).

An NVIDIA slide showing off the more consistent CSAA.

Combined with PureVideo HD (offering spatial temporal deinterlacing and Inverse Telecine for VC-1 and H.264 sources, extra noise reduction and edge enhancement), NVIDIA are for the first time touting an IQ victory over ATI, claiming that thanks to these improvements it score higher in HQV than the X1950XTX.

Implicit in all of this is DX10. You'll need Vista to take advantage of it (in fact it'll probably have gone RTM by the time you read this), and indications are the transition to DX10 will be quicker than that of DX8 to 9. Games so far indicating support are Crisis, Flight Simulator X, Quake Wars: Enemy Territory and interestingly, Half Life 2: Episode 2. You can certainly tell the difference in Crysis between DX9 and DX10.

Hellgate London and World in Conflict will support Quantum Effects.