nVIDIA's Initial Approach to PCI-Express: Building a Bridge from AGP
nVIDIA has long been the leading seller of high-performance graphics chips (GPUs) for various versions of the AGP bus. Despite the introduction of PCI-Express chipsets and motherboards in mid-2004, AGP is still the mainstream high-speed graphics choice for most users. To move into the PCI-Express marketplace, nVIDIA had a choice: develop brand-new PCI-Express GPUs, or figure out a way to adapt its high-performance AGP GPUs such as the GeForce FX to work with PCI-Express. For its initial releases, nVIDIA has chosen option #2, using a high-speed interconnect (HSI) chip to act as a bridge between the AGP bus support in its GPUs and the PCI-Express bus connecting the cards to the motherboard.
Generally speaking, I'm not a fan of bridged solutions (such as ATA to Serial ATA drive interfaces) because they tend to be slower than a native implementation of a particular technology. However, if a bridged solution runs as fast as a native solution, I don't see a problem with it. According to Tom's Hardware's article Future Promise for Graphics: PCI-Express http://graphics.tomshardware.com/graphic/20040310/index.html, the bandwidth and latencies of the HSI chip are fast enough to avoid bottlenecks.
By using the HSI chip with its existing line of GeForce FX GPUs, nVIDIA was able to re-use its high-performance AGP GPUs with the new chip architecture.