QPI

Up until Nehalem, Intel has stuck to the time-tested processor methodology of having the CPU communicate with the rest of the system via the single, bi-directional interface we all know and love - the Front-Side Bus (FSB). However, the FSB has pretty much reached its limit and as much as Intel could try to tweak its CPUs to offset the problems a better solution was needed.

Intel's answer to the problem is twofold; adopt an integrated memory controller (IMC) and introduce a new interconnect called QPI to cater for everything else the FSB did. One primary advantage of QPI versus the FSB is that QPI is a point-to-point interface, so there's no competition for bandwidth. If the phrase "Intel-flavoured HyperTransport" springs to mind" then you're on the right lines.
/94/d314db/b28c/9158-nehalem22.jpg
Each QuickPath Interconnect port offers 25.6GB/s of bandwidth at 6.4 Giga-Transfers a second with 20 lanes in each direction. As if that wasn't enough, the implementation is scalable so, should more bandwidth ever be needed, Intel can adapt QPI to fit those needs.

Not only does QPI allow cores on the Nehalem die to talk to each other and the IMC, but it also serves as the communications highway between the CPU and motherboard chipset. This QPI link will allow the processor to talk to the PCI Express, USB and SATA controllers, which will continue to be managed by a chipset on Nehalem platforms, as it needs to.
/94/5f4cfb/0094/9158-nehalem23.jpg
High end Nehalem CPUs will sport a pair of QPI links, while mainstream chips will have only one. The reason for the second link on higher-end systems is that one QPI link will go to the chipset, while the other will be used for inter-CPU communication in multi-socket configurations.

The only thing QPI doesn't serve as the CPUs conduit for is RAM. For that Nehalem has an IMC. Thanks primarily to AMD we already know the advantages of an integrated memory controller. Fundamentally the biggest benefits are the huge increases in available bandwidth and reduction in memory latency gained by letting the CPU talk to RAM directly, although that's not the full story.

comments powered by Disqus