R423 and PCI Express Benefits
ATI's Radeon X800 graphics processors were well documented recently, the new R420 GPU enabling a new leap in general 3D performance, without really pushing the boat out in terms of new features. While it's tempting to stuff this page full of information on R420, the GPU that powers the X800 series of cards, I'll save you the suffering. If you want to read about R420 as an architecture, the links above will sort you out.That out of the way, it's really easy to talk about R423 as a GPU, since it's nothing more than R420 with the AGP interface removed and a PCI Express 16X interface augmented in its place.
Built on the same 0.13µm low-K process at TMSC, with the same basic 16 pixel pipes, 6 vertex shader units and 256-bit memory interface, nothing else changes bar the physical interface. However, it's the basic interface change that's the highlight, the PCI-SIG members not creating the spec and interface because they had nothing better to do.
PCI Express Benefits
I'll quickly sum up the benefits of PCI Express here, in terms of graphics connectivity on a 16X lane, so that the chief advantages of the interface can be referred to for R423 and also for the other GPUs mentioned in this article.While the most obvious benefit of a 16X PCI Express lane is its bandwidth increase over AGP8X, it's the way that bandwidth is able to be used that holds the key to why the transition is being made in the first place. AGP8X specs up to 2.1GB/sec from its 4-byte (32-bit) bus. Based on the venerable PCI bus, 8X means 8 data samples per clock, ticking at 66(.66)MHz for an effective 533MHz.
PCI Express works a bit differently. The specification defines a basic 1-byte (8-bit) 'lane'. Each lane is a serially connected I/O interface with one upstream path and one downstream. A 16X implementation means 16 1-byte lanes for 16 bytes in total, with independant up and downstream operation.
With the 16X specification offering 4.2GB/sec in both directions, on a 16-byte bus that means a base clock of 266(.66)MHz.
The key to PCI Express's benefits over AGP8X aren't just the doubling of available bandwidth, it's the fact that the bandwidth increase is available in both directions asynchronously, to and from the GPU and the PCI Express host. Current AGP8X is 2.1GB/sec in one direction, from the host to the GPU. GPU to host reads and writes currently hover around 300MB/sec, far below the upstream bandwidth and a giant leap below what's available with PCI Express.
Having 4GB/sec downstream at the behest of the GPU is the main bonus of 16X PCI Express over AGP8X, the increased upload bandwidth secondary.
Electrical benefits run mainly to a larger current draw over the socket, reducing the need for extra power connectors on a lot of boards.
