The new Socket LGA1155, common to both mainstream desktop Sandy Bridge Core i3/i5/i7 chips and the entry level Xeon Sandy Bridge parts, will support CPUs in two different configurations for PCIe I/O: the 16-lane desktop part, and the 20-lane server and workstation part, all of course at PCIe v2 speed.
Otherwise, the desktop and enterprise parts are identical, including up to four cores and 8MB cache, and speeds up to 3.4GHz for the parts with GPU turned on, and 3.5GHz for the parts with the disabled GPU. The fine grained Turbo capability gives them another up to 400MHz headroom when all cores are used, with appropriate power and thermal solutions. These sockets, by now well known to the community, are the only ones with the built-in GPU.
Then we come to the Socket LGA1356, a direct replacement for the current Socket LGA1366. The parts here are 6-core and 8-core Sandy Bridge single-socket and dual-socket capable but midrange positioned Sandy Bridge Xeon - and, ultimately, Core i7 - parts with up to 20MB of L3 cache, three DDR3-1600 memory channels just like the existing LGA1366 Westmeres with one memory speed grade higher, and 24 PCIe v3 lanes on-chip. The single external QPI v2 link runs at up to 8 gigatransfers/sec, or 32GB/sec bidirectional bandwidth, a 25 per cent speed up over the current generation, but also feeding a third more cores on each socket.
The highest speed 8-core CPUs with up to 150W TDP should, however, be reserved for the high-end Socket LGA2011. With more power and ground lines to support 40 PCIe v3 lanes and four DDR-1600 memory channels per socket, as well as dual QPI 8 gigatransfers/sec links, the 8-core, 20MB L3 cache Sandy Bridge-based Xeons should have sufficient system bandwidth to feed even the highest workloads. Not to mention enough PCIe bandwidth for two dual-GPU cards with extra lanes still free for a, say, 5GB/sec PCIe high-speed SSD or Infiniband interconnect.
And, when you add the same resource on the second CPU, it becomes possible to fully feed an 8 GPU system out of a single two processor workstation. And yes, you could even do a quad-socket monster here, if you're using the EX parts, I assume.