What is PCI-Express

Now all things considered, it is probably a good thing that the PCI bus has remained essentially unchanged since its inception. For one thing, it works.

It has provided a stable and flexible platform for hardware and software developers to build on for almost a decade. Anyone who remembers the days before Windows 95 and 'Plug and Play' devices will understand why computers have become so much more common since then.

Now there are three other PCI specifications in existence, all designed to increase the amount of available bandwidth. These are 66MHz PCI, PCI-X at 64bit/133MHz, and the soon to be introduced PCI-X 2.0.

The trouble is, while these technologies have, or soon will find a permanent home in the server market, the complexities and extra costs they introduce to motherboard manufacturing mean that they will be virtually unknown at the desktop level. PCI-X, for example, requires a controller for every slot and that is just too expensive. The solution to this is being backed by everyone's favorite processor manufacturer, Intel.

Intel, in partnership with several other companies which include the likes of IBM, Dell, Compaq, HP and Microsoft, recently introduced what will be the new standard for PC I/O in the years to come; PCI Express (formerly 3GIO or 3rd Generation I/O) is the name which it was given by PCI-SIG, the committee responsible for overseeing the PCI interface standard.

Approved as a standard on April 17 2002, PCI Express is an evolutionary upgrade to the existing PCI bus. It co-exists with the now 'legacy' PCI bus, but in terms of form, it is something completely new.

What does PCI Express mean to your PC?

Like AMD's Hypertransport technology, PCI Express is a 2-way, serial connection that carries data in packets, similar to the way it is transferred over Ethernet connections.

The PCI Express bus is no longer a single parallel data bus through which all data is routed at a set rate. Rather, an assembly of serial, point-to-point wired, individually clocked �lanes� each consisting of two pairs of data lines carries data upstream and downstream. As the technology goes to market, each of these lanes is capable of a 2.5Gb/s data rate in each direction. The overall sustained transfer rate roughly equals 250MB/s.

PCI Express was designed with future upgrades in mind, and should see increased results with later enhancements. Where the real performance benefit comes in is when more than one lane is added to a given point-to-point route. Lanes can be stacked together to increase the amount of bandwidth available to specific areas of the I/O system, such as the video card slot, for example.

This technology has a couple of immediate benefits. First of all, the amount of trace routes on any given motherboard areconsiderably reduced by the adoption of point-to-point circuitry.

Using 4 routes instead of 32 for a basic connection, you can see how motherboard manufacturers might well be enthusiastic about this technology, as this would significantly cut the costs of producing motherboards in the long term. Secondly, since each lane is exclusively used for communication between two points, there is no sharing of the available bandwidth.

The objective then, was to create an I/O bus technology that will have the staying power that 32-bit PCI slots have enjoyed, a decade or more into the future. It seems like it has been a success thus far.

Upcoming technologies such as Intel's multi-Gigabit Ethernet 'Dual Northway" chipset, will inevitably require more bandwidth than PCI can currently offer, so its phasing out in the server market seems inevitable.

The biggest impact that PCI Express has made on the desktop market is with the PCIe x16 graphics slot. Found in the latest Intel and AMD-based chipsets, This implementation of the technology has virtually replaced AGP 8x as the platform of choice for graphics card manufacturers. All new graphics cores from both nVidia and ATI support PCI Express x16 by default, though some have been ported to AGP 8x.

The 164-pin X16 slot provides a total usable bandwidth of around 4GB/s in each direction, double the 2.1GB/s bandwidth that the 8x AGP spec boasts. The X16 specification is pretty similar to AGP in terms of connector card size, though entirely non-compatible, which gives it a boost in terms of familiarity and ease of installation.

nVidia has succeeded in creating dual-SLI graphics technology using the PCIe x16 slot and its new nForce 4 chipset . This technology allows two PCIe x16 slots to split 16 lanes of PCI Express into two sets of eight and run two identical video cards in tandem. This leads to often dramatic boosts in 3D gaming performance.

Videocard Capabilities

Currently, PCI Express X16 provides up to 75W of power for video cards, as opposed to the current 25W/42W allowed by AGP8X. Most modern cards already consume more than this amount, as evidenced by the extra power connectors that appear on today's AGP high-end video cards. This explains why most modern midrange PCI-Express videocards do not require extra power connectors, while their AGP counterparts do.

Given that both ATI and Nvidia have virtually switched over to PCI Express and that major graphics companies from 3Dlabs to XGI have already announced support for the architecture, it looks like AGP's days are inevitably drawing to a close. To further depreciate the value of your very expensive AGP cards, PCI Express X16 and 8x AGP slots cannot coexist properly on the same motherboard. This fact has not stopped various motherboard manufacturers from releasing ersatz PCIe/AGP solutions that use the PCI bus or part of the PCI Express bus. While this provides a better incentive to upgrade for AGP card users, it should be noted that these solutions do not provide the full bandwidth of a real AGP8x slot.

What makes up the PCI Express bus?

The PCI Express bus is composed of multiple lanes of point-to-point wired copper interconnects serving all the components that used to just drop data into the older PCI bus. To make things even more flexible, a switch has been added to the architecture.

Working in principle just like an Ethernet switch, this sits between the PCI Express devices connected to the board and the rest of the I/O system. This will enable newer devices, for example PCI Express connected on board Gigabit controllers or 10/100 gigabit network cards, to communicate with each other directly if necessary. By adding in this shortcut the data need not go through the chipset to reach an adjacent device.

Compatibility between PCI 2.2 and PCI Express

As mentioned previously, the PCI Express bus is software compatible with PCI 2.2 compliant devices. As Intel stated, all current operating systems will support PCI 2.2 compatible interface cards in PCI Express systems using the existing PCI drivers. This is because the legacy bus is still present, just in a reduced form.

This was an essential requirement if PCI Express technology was to gain widespread acceptance. Forced obsolescence does not tend to go over well with computer users.

PCI Express chipsets and connectors

The current standard PCI Express implementation, as seen in chipsets from Intel, nVidia and VIA, allows for 20 lanes of PCI Express in total. 16 of these are used for the PCIe x16 graphics slot and four more are distributed between any combination of PCIe x1, x2 or x4 slots.

Current PCI Express capable chipsets include Intel's 915, 925 and upcoming 955, nVidia's Nforce 4 and nForce 4 SLI Intel Edition and VIA's K8T890, among others. Both AMD and Intel processors are now well-supported with PCI Express, though Intel chips had the technology well before AMD chips realized it with the introduction of the nForce 4.

Other than video cards, there are few PCI Express-based peripherals on the market, but with nVidia and ATI's products spearheading a huge adoption of PCI Express capable motherboards, it should not be long until peripherals based on the technology are currently available.

It was predicted that Gigabit Ethernet adaptors would be the first technology to be commonly available using PCI Express besides video cards, but since most PCI Express capable motherboards now integrate Gigabit adaptors, there is little demand for them in the desktop market.

A PCI Express 1X connector is ideal for gigabit Ethernet cards, which have enough throughput to potentially cripple a standard PCI 2.2 bus. As found in most modern PCI Express motherboards, the Gigabit Ethernet controller is soldered directly to the motherboard and runs through the PCIe bus.

The BTX form factor (announced at IDF 2003) is intended as the replacement for the current ATX motherboard standard, and included in its specifications are an X16 slot instead of an AGP connector for graphics, and two 1X PCI Express connectors in place of several of the PCI slots.

In the picture to the left, the PCI Express slots are coloured blue. The BTX formfactor is beyond the scope of this article, so for more information about it, look here.

PCI Express X1, X2, X4, X8, X16

PCI Express connectors are similar in appearance and connection method to 32-bit PCI slots. PCI Express 1X slots are about the size of current modem riser slots (about 1" long), while the X16 interface (164-pins) for graphics is very similar in appearance to the standard AGP port. The flexibility to adapt to PCI express devices of different bandwidths is built into the midrange X4 and X8 slots that we have seen.

For the portable market, a new standard called the 'ExpressCard' has been created by the PCMCIA group in cooperation with the development of PCI Express by the PCI-SIG group. Many of the companies involved in the work on PCI Express are members of both trade associations, so it can be considered a joint effort.

The ExpressCard is clearly meant to be a technology that bridges the gap between current notebook designs and future PCI Express enabled ones.

The interface and form-factor (similar in appearance to the current PC Card form factor for laptops) supports both USB2.0 and PCI-Express, allowing manufacturers to implement one or both without worrying about being ahead of or behind the times.

Intel has used ExpressCard technology as an integral part of their 'Marble Falls' small form factor PC technology demonstration. ExpressCard slots will not be compatible with previous PC Card peripherals.

Adding to the potential flexibility of the PCI Express system, is the capability of external connections. The same 200MB/s point-to-point connections could be extended over cable outside the box to a distance of approximately 5 meters. Add this to the fact that designated PCI Express connectors will be hot-pluggable, and you have the potential for some very interesting uses of the technology.

External PCI Express devices road mapped

Among other things, Intel has suggested a topology for the future that would separate the computing unit from the monitor, interface devices and external drives using PCI Express cable connections as a go-between.

A concept PC like that would enable complete modularity in servers and even desktop systems if manufacturers are willing to accept it. Of course, this is significantly beyond the current plans for PCI Express, but if the technology is capable of supporting more user friendly expressions of computer form, they will probably be developed.

PCI was developed in part to make computer configuration easier to handle, so there is no reason why PCI Express should not take this to a new level. External video cards, anyone? Speaking from past experiences, I don�t know how many times external components have been mentioned as a �wave of the future� type of thing; how about an entirely component based PC like a stereo system?

It is unlikely that the makers of Firewire and USB 2.0 devices have much to worry about in the short term, despite the apparent ability of the PCI Express technology to carry out their roles without even requiring a bridge to connect to the rest of the I/O system. USB and Firewire devices are already very common, and by the time external devices supporting the PCI Express specification become available in a year or two, they should have become so prevalent in the market that manufacturers will have to continue support for them for the same reasons that PCI Express was designed to coexist with PCI 2.2.

Also, the ExpressCard format explicitly combines USB 2.0 and PCI Express in a single slot for portable and small form factor devices. This is both an indication that PCI Express is not necessarily intended to supplant USB anytime soon, and a concession to the continued usefulness of USB technology in the brave new world of sometime next year.

PCI Express does not compete with AMD's Hypertransport technology, but rather coexists with it in order to move data more efficiently in chipsets like the nForce 4 and K8T890.

Likewise, PCI Express is intended to serve as the underlying data-transport technology for I/O methods like serial ATA and SCSI, not to replace them. PCI Express is going to replace AGP as the conduit between the graphics chipset and the processor and memory however.

As you can see, the modularity of the PCI Express system should enable it to carry out many different roles in the computer industry. As manufacturers will have the capability of deciding how much bandwidth to assign the needs of a specific market. Commenting on server technology is slightly beyond the scope of this article, but it is easy to see how PCI Express technology could work well in this high-data area.

The PCI Express architecture is not intended to replace current CPU/memory bus technology, but rather to work with it by providing more available bandwidth and flexibility for the rest of the I/O system.

While it was originally believed that the PCI Express developers group might be looking at ways to standardize the CPU and memory buses within the system as well this is not part of the final approved specifications.

Industry heavyweights ATI, Nvidia, and even AMD have already given PCI Express their support for years to come. With endorsements like that it seems inevitable that the PCI Express I/O system will be the basis for the forseeable future.