PCIe: A Brief Technology Primer On PCI Express
Myth: A 16-lane PCIe slot always enables 16 lanes
As Confusing As It May Sound, A x16 PCIe Slot Is Different Than A 16-Lane Configuration
PCI Express as a standard can be confusing. First consider the physical slots. They’re commonly referred to as x1, x4, x8 or x16. This nomenclature makes reference to the slot’s physical size and the maximum number of PCIe lanes that a card inserted into that slot can access.
Smaller cards fit into longer slots (for instance, a x8 or x1 card into a x16 slot), but larger cards do not fit into shorter slots (a x16 card into a x1 slot, for example). With rare exceptions, graphics cards are almost universally designed to communicate over a 16-lane link, thus requiring a corresponding slot. In theory, the PCIe specification allows for up to 32 lanes per slot, though we’ve never seen anything longer than x16.
PCIe 32x slots/cards?
A little-known fact is that, in theory, the PCIe specification allows for up to 32 lanes to be used with any given slot. We have never seen slots wider than 16x however ... the physical size of 32-lane slots would be really big though and, likely, close to impossible to implement on ATX form factor motherboards (as they would create a routing nightmare).
So, let's say that you have your 16-lane graphics card fitted nicely into a x16 slot. Does that mean it’s transferring data across a 16-lane link? Maybe, but maybe not. The number of lanes actively associated with each slot depends on the host architecture (Haswell or Haswell-E, for example), the presence of bridge chips (such as PLX’s PEX 8747) and on the number of cards installed in surrounding slots (hardware strapping on most motherboards dynamically reconfigures lanes based on utilization). Thus, a x16 slot might operate with 16, eight, four, or even one lane active. The only way to tell for sure is through a tool like GPU-Z, though even that can be unreliable.
Without going into excessive detail, a major difference between Intel's "premium" LGA 2011 interface (supporting Ivy Bridge-E-based CPUs like the Core i7-4820K, -4930K and 4960X) and Intel's "mainstream" LGA 1150 interface (supporting Haswell-based processors like the Core i7-4770K), is that the higher-end platform exposes up to 40 lanes of third-gen PCIe natively. That’s enough for two cards in x16/x16, three cards in x16/x16/x8 or four cards in x16/x8/x8/x8 configurations). Meanwhile, LGA 1150 is limited to 16 lanes, giving you enough connectivity for one card at x16, two cards in x8/x8, or three cards at x8/x4/x4 in CrossFire (Nvidia does not permit SLI using four-lane links). There are exceptions of course. Intel’s Core i7-5820K drops into 2011-pin sockets, but is deliberately limited to 28-lane PCIe controllers.
To work around the limited PCIe connectivity available on LGA 1150 platforms, hardware manufacturers use PCIe "bridge chips" that operate like switches, enabling access to a greater number of PCIe lanes. The most famous of those chips, PLX’s PEX 8747, is a PCIe Gen 3 switch, and it isn’t cheap. Mouser has them at roughly $100 each for low volumes. OEMs likely negotiate lower prices on greater quantities, but that remains an expensive component. This chip alone contributes to the jump from ~$200 to ~$300+ between mainstream and enthusiast LGA 1150 motherboards.
What does all of that translate to from a bandwidth perspective, though?
As you can see, each PCIe generation essentially doubles the bandwidth of the version before. Four lanes of PCIe 3.0 are roughly equivalent to eight lanes of PCIe 2.0, which in turn are roughly equivalent to 16 lanes of first-gen PCIe.
The fourth generation, scheduled for 2015, will be backwards compatible with today’s technology. You can expect PCIe 3.0 (and 2.0) graphics cards to remain current for quite a while, at least from an interface standpoint. If you're already curious about PCIe 4.0, read the PCIe 4.0 FAQ directly from PCI-SIG.
We still haven’t answered the question of how much bandwidth you need, though.