The new Input/Output (I/O) standard, PCI Express, changes all that. Appearing in systems starting in 2004, PCI Express is technically not a new generation of PCI architecture, but an architectural leap. It keeps the core of PCI's software infrastructure,

ATI Technologies Inc. said the company had shipped a million PCI Express graphics cards so far, in line with what the company sees as a rapid transition to the new interface standard. – ExtremeTech*, August 13, 2004

but completely replaces the hardware infrastructure with a radically new forward-looking architecture that rockets I/O back into the express lane of performance.

PCI Express is not only designed to replace the PCI bus for devices such as modem and network cards, but also the Advanced Graphics Port (AGP) used for desktop graphics cards since 1997. Unlike PCI, its parallel predecessor, PCI Express is a serial point-to-point interconnect capable of extremely high-bandwidth transfers. Performance ranges from 250 megabytes per second (MBps) for a single "lane" implementation and up to 4 gigabytes (GBps) for the 16 "lane" implementation likely to dominate the AGP-replacement market. Better yet, there's room to grow and for even faster I/O.

As fast as this is, it doesn't spell the end for PCI. Like the 16-bit ISA (Industry Standard Architecture) bus that PCI replaced, PCI will likely be around for some time, used in implementations that don't require the performance gains of PCI Express. In fact, for the time being, many desktops systems will probably offer PCI expansion slots right alongside one or more PC Express slots.

PCI was originally developed through the efforts of Intel and other industry leaders in response to the realization that ISA, a bus design developed in 1984, was becoming a bottleneck within the computer. Intel wanted a high-performing bus and thought the best solution would be to pursue an industry standard. Having a standard would prevent the complication of competing buses and consumer confusion. A technology under development in Intel research and development seemed the perfect solution, offering both superior performance and plug-and-play capabilities. Intel spurred the industry by:

"Intel is in the business of providing the engine for the PC, just like Honda is in the business of providing the engine for the automobile. That [PC] engine is doubling in capacity every 18 to 24 months — that's Moore's Law. What we really want is to ensure the rest of the platform goes with it. This means that if the engine gets better, the tires get better, the chassis gets better, the roads get better, and you get better gas mileage ... we want the platform, which is everything around the microprocessor, to be keeping pace and improving and scaling, so the microprocessor can deliver its potential." – Bala Cadambi, PCI Express initiative manager at Intel, when he was working on PCI

• Forming a PCI special interest group (PCI-SIG) with others in the industry.
• Contributing the technology from Intel research and development.
• Encouraging an open industry specification.
• Implementing the new specification in Intel product groups, ensuring ease of implementation and availability.

The result was PCI — one of the most (if not the most) successful chip and board interconnect technologies in history. PCI provided a tenfold performance gain over ISA, and its plug-and-play capabilities made it easier for users to add cards to their computers. Developed initially for personal computer applications, PCI is used today in almost every computing platform.

As fast as technology has evolved and advanced over the last 12 years,

One type of forum that Intel relies on to refine a new standard and help companies develop prototypes is a compliance workshop popularly known as a "plugfest." Often hundreds of companies attend — sending two or three of their best engineers — to make sure prototypes work with a new standard like PCI or USB. These plugfests help create legitimacy and support for a new standard.

it's surprising how long PCI has lasted. PCI's longevity can be attributed to the PCI-SIG and two characteristics of PCI. The first is that PCI is processor-agnostic — both to frequency and voltage — so it could function in the desktop, mobile and server markets with little or no change. It wasn't wired into a specific processor, so manufacturers could standardize their I/O across multiple product groups and generations. That also meant computer manufacturers wouldn't have to redesign their PC architecture every time Intel came out with a new chip. The second characteristic is that PCI is flexible in its ability to support multiple form factors. PCI-SIG members were able to define connectors, add-in cards and I/O brackets to standardize the I/O back panel and form factors for the server, desktop and mobile markets. This standardization made the distribution of PCI-based add-in cards and form factor-based computer chassis possible through the consumer channel and in sufficient volumes to meet consumer price targets.

During the Spring 2001 Intel Developer Forum (IDF), Intel presented its case for replacing the PCI bus and the various internal chip interconnects. Intel suggested a new Third Generation I/O (3GIO) technology and a PCI-SIG working group was formed.

As with PCI, Intel research and development contributed key technology and technical expertise.

"...don't think that PCI will go away anytime soon after PCI Express is introduced. Recall that the 16-bit ISA bus has been with us since late 1984, with ISA signaling still present in Super I/O legacy devices. We're finally at the point where most new systems do not include ISA slots." – Nick Stam, ExtremeTech*, September 9, 2002

Work progressed quickly and within six months, the PCI-SIG approved the Arapahoe standard — a second codename for 3GIO. Momentum rapidly picked up as other companies supported the new open industry specification and joined the group. By April 2002, the draft specification was finished and the name PCI Express was coined.

To encourage implementation and ensure availability, Intel began implementing the new specification across its product groups. Chipsets and graphics chips based on the standard began appearing in early 2004. The first systems rolled out in mid-2004.

Parallel-bus architectures like PCI have a fixed, one-way maximum bandwidth that is shared. That means bus frequency can't be increased to match increases in processor frequency. It also means as boards are added to a PCI bus system, the bandwidth available to each function is reduced. Contention for the bus, where a higher priority transfer can suspend lower priority data, becomes a real-time problem.

PCI Express, on the other hand, is a bidirectional, point-to-point serial protocol. Data can flow upstream and downstream simultaneously. What's more, it's scalable. To keep up with increases in processors speeds, I/O designers simply add more "two-way lanes." And since it's a series of point-to-point, individually clocked, wired "lanes," there's no contention for the bus. Each device has its own "highway." Each lane is capable of a 256-MBps data rate in each direction. Add more lines for a faster data rate. The PCI Express specification supports widths of 1, 2, 4, 8 16 and 32 lanes. At 32X, that's a bandwidth of 8 GBps per stream and a total of 16 GBps for both streams.

PCI Express has many other advantages as well. It supports important features such as power management, hot-swappable devices, and the ability to handle both host-directed and peer-to-peer data transfers. This last capability means PCI Express can emulate a network environment — it can send data between two compatible points without routing it back and forth through the host chip.

PCI Express also simplifies board design. Its single-wire, serial technology eliminates the strict wire count or length requirements of parallel bus architectures. This should help keep board implementation costs at or below those of PCI.

The PCI-SIG is initially positioning PCI Express as a general-purpose I/O technology for desktops, workstations and servers that will also speed internal chip-to-chip and graphics adapters.

The first PCI Express add-in devices for desktops will be graphics cards. That's because like PCI, the AGP standard is running out of gas. PCI Express at 16X is about twice as fast as the current generation AGP (8X). PCI Express also provides a full duplex interface, while AGP is only a half-duplex interface. Data can only be going in one direction at a time, not both directions simultaneously like PCI Express.

One leading graphics card manufacturer estimates that, by the end of 2004, half of the PCs sold will contain a PCI Express graphics chip. – ExtremeTech*, August 13, 2004

Yet another advantage is that PCI Express provides additional power to the cards — almost triple the maximum power draw. High bandwidth and more power are a tough combination to beat.

PCI Express won't just be making an impact on desktops and servers, either. The PCMCIA and CardBus standards will be replaced by ExpressCard. This card combines PCI Express and USB 2.0 to enable a higher bandwidth, modular expansion architecture for notebook PCs.

One of the more intriguing aspects of PCI Express is the flexibility it gives manufacturers in redesigning the traditional beige box. With fewer circuits to design onto the motherboard, a PCI Express slot can be just 3 inches long and 1.3 inches wide. This will enable vendors to experiment with compact desktop designs that don't sacrifice expandability.

The modularity of PCI Express could even lead to topologies that separate the CPU unit from the rest of the computer. Think of an entirely component-based PC that's analogous to a home stereo that has a separate amplifier, CD player, tape player and FM radio tuner. Using a PCI Express cable, PCs could have external DVD and hard drives. The advantage would be an ability to replace any of the units when something better was available without having to replace the entire computer.