In somewhat surprising fashion, NVIDIA has chosen to approach the transition to PCI-Express with a unique new conversion method called High-Speed Interconnect (HSI) technology. Without question, NVIDIA has the burden of proving that their new HSI technology is able to counter the pitfalls one would associate with a generic bridged solution. In its simplest form, the image below represents the basic configuration of a HSI chipset which connects an AGP GPU to a PCI-Express host. Keeping this setup in mind, we will compare NVIDIA’s HSI model to a conventional bridged solution to uncover the pros and cons to this new technology.
The goal of this new HSI interface is to allow any GPU in NVIDIA’s product line to be used with either an AGP slot or a PCI-Express slot. Although the HSI chip will allow NVIDIA to gradually increase its penetration into PCI-Express as it is adopted, the real benefit will come down the road a bit. Once PCI-Express cards become the norm, this same chip will allow people with AGP motherboards to still purchase the latest graphics cards as well.
The image below represents a generic bridged solution between the GPU and the PCI Express host. Because the components of this system have all been designed around a straight conversion with no design considerations for optimization, there will be significant bandwidth limitations and latencies. In short, this generic solution would not be able to compete with a native PCI-Express solution.
Easily the most critical point of concern with this new technology is the issue of bandwidth. With AGP8X being limited to ‘only’ 2.1GB/s, the 4GB/s offered by PCI-Express seems to make this debate a non-issue. However, we must realize that we are not bound to a strict AGP8X standard as ATI is leading you to believe. Rather, as long as the GPU can support it the HSI allows for downstream speeds equivalent to AGP16X. As seen in the image below, this equates to an overall bandwidth of 4.2GB/s which is on par with PCI-Express. In terms of upstream bandwidth, the HSI setup is limited to 1GB/s bandwidth versus the 4GB/s bandwidth of PCI-Express. However, we must realize that games and common applications primarily use downstream-bandwidth so the adverse effects of this for the foreseeable future are limited.
What we have seen thus far is that much of ATI’s criticism of NVIDIA’s approach is based upon the use of a generic bridged solution and not NVIDIA’s HSI solution. Specifically, these higher latency claims target an architecture with a 32B request size. The 32B request size not only amplifies the effect of latency within the system, it can also constrain the overall effective bandwidth of the design. As such, it is important to note than NVIDIA’s architecture supports a 64B request size which minimizes the effect of any introduced latencies dramatically and allows for the maximum possible bandwidth.