The Little Bridge That Could
When the PCI Express x16 graphics interface was first settled on being the next generation graphics I/O (input/output), NVIDIA had already identified that the existing AGP 8x interface would still linger around for a more couple years in a vast majority of systems. Having experienced the PCI to AGP interface change for graphics cards, they had a good gauge of the timeline involved before the newer interface progresses to be the de facto standard for graphics. Unlike the AGP interface that was an extension and evolution of the PCI interface, PCI Express (PCIe) is radically different because it is a very fast and narrow serial point-to-point topology versus the old parallel PCI architecture that had difficulty in future performance scaling. With a new communications topology, signaling and the way data transfers were scheduled too differed. Instead of increasing the number of GPU SKUs required to natively support both PCI Express and AGP graphics cards, NVIDIA had poured much resources to invent their High Speed Interconnect (HSI) chip.
This is the NVIDIA High Speed Interconnect (HSI) chip situated immediately below the GPU. Although it has a huge packaging (22mm x 29mm) for its many pin outs, the die itself is relatively puny at only 27mm².
This HSI bridge chip allows a native NVIDIA AGP GPU to communicate across the PCI Express bus and vice versa where a native PCI Express GPU can utilize the traditional AGP bus. The ability to convert the bus interface used for communication via this bridge chip opened a whole lot of possibilities and savings in both time and cost for NVIDIA in the long run. For example, when the first PCI Express enabled motherboards shipped, NVIDIA had all the required GPUs (which were only compliant with the AGP bus) at its disposal and only needed a PCB redesign and incorporating the HSI bridge chip successfully. This approach is vastly speedier (and cheaper) than redesigning and re-fabricating the GPU to be natively compliant with the bus interface intended for use. More so because the PCI Express interface has yet to find acceptance in the general consumer market. Therefore, the reuse of existing GPUs in conjunction with the HSI bridge seemed quite logical and allowed faster time-to-market penetration for PCIe graphics cards as well as better inventory management. Taking the early NVIDIA GeForce PCX 5900 for illustration, it was actually using a GeForce 5900 XT GPU paired with a HSI bridge on a redesigned PCB.
The GeForce 6600 GT on the other hand, was designed specifically for the PCI Express market, but with NVIDIA's development of the HSI chip, it allowed them to deploy this PCIe GPU part in the AGP graphics card market as well.