Inside the GeForce 7800 GTX

High Dynamic Range Rendering

GeForce 6800 Ultra was the first time that we had seen a realisation of High-Dynamic Range Rendering – it looked fantastic, but there was one problem, in that the performance generally sucked. There was just too much being asked of the GPU at any one time for it to be useable at any kind of decent resolution.

GeForce 7800 GTX has been optimised to work well with High-Dynamic Range rendering, meaning that there should be a much-needed kick up the arse for the previous performance-related issues that we came across when we first played around with HDR in Far Cry all that time ago. Even when running two GeForce 6800 Ultra's in SLI, we struggled to get the necessary performance to deliver a smooth frame rate at a decent resolution.

The optimisations inside G70 include improved texture caching and computation subsystems that have been designed to be faster than NV4x's HDR performance per pipeline at the same clock speed, even when not limited by shader performance. Along with that, the high-precision shader data has also been optimised in the pixel shader to further improve HDR performance.

Bear in mind that you cannot use Anti-Aliasing when using HDR, as both techniques need to be the final operation before each pixel's colour value is stored in memory. Thus, the quest for higher resolutions to eliminate as many aliased textures as possible is a requirement.

We'll have a closer look at the performance that can be expected from games making use of HDR in the second part of this series of articles. Image quality hasn't changed at all, but speed has improved drastically, meaning that you can increase the resolution to a level that looks good enough to complement the improved lighting and remove some of the aliased textures.

Power

You would assume that a larger, faster, more complex graphics-processor would require more power than the previous generation - if you did, you'd be wrong. Thanks to the smaller manufacturing process that has been used for G70, GeForce 7800 GTX uses less power than GeForce 6800 Ultra – consuming less than 100W under heavy load.

NVIDIA have also implemented something similar to PowerMizer – as seen on NVIDIA's mobile parts. In a nutshell, this means that power scales with activity, reducing power when the GPU is idle – for the full low-down on PowerMizer, you can read our GeForce Go 6600 preview.

After the numerous power problems that people had with both SLI and single GPU configurations, NVIDIA have been clear to state the minimum requirements for the power supply that is destined to power GeForce 7800 GTX. For a single GPU configuration, NVIDIA recommend a minimum 350W PSU and a 500W PSU for those of you wanting to grab a pair for some SLI love. (What a nice pair they would make too – Ed.)

Initial Thoughts...

Without mentioning too much about the video cards' physical appearance and focussing on the architecture behind the GeForce 7800 GTX, it would seem that there are some smart moves that will enable them to continually scale shader performance without having to drastically increase the number of pipelines and increase the overall size of the graphics-processing unit to levels that will cause problems with yield and cooling.

The throughput of each pixel shader unit is quite outstanding when you consider how efficient NV40 already was – developers loved developing games on NV40, and G70 looks like it will reap the benefits of the last 18 months of game development thanks to its massive pixel throughput. NV40 paved the way for more efficient graphics processors, and the logic of doing more work on each pixel per clock cycle.

The NV43-esq pixel output engine design is a smart move too – I was speaking to one of NVIDIA's dev-tech engineers last December regarding the way that NV43 was designed, and how the NVIDIA ethos had changed a little from previous years. The focus is clearly on making the most out of the silicon that they've budgeted for, and wasting as little of the GPU as possible. We mentioned it earlier, and we'll mention it again – it seems that idle silicon is wasted silicon in NVIDIA's eyes.

Providing parallelism and width keeps working in 3D, we find it hard to believe that GPU's are going to hit the same performance wall that CPU's have hit in recent times. With the way that things currently stand, a focus towards greater parallelism and doing more work on each pixel in every clock cycle, along with improvements in power saving technology sums up the way forwards for graphics and NVIDIA seem to have adopted that ethos, at least theoretically.

Don't forget to check out the second article that focuses on what you can expect GeForce 7800 GTX to deliver in terms of real-world gaming performance in both single card and SLI mode.