Remodeling Input/Output

Platform change is a regular occurrence in the world of PC hardware, and we’re bracing for a big wave to hit in 2004. Intel and AMD are expected to make a transition to new CPU packages, DDR-2 memory may or may not become the de facto performance standard, and Serial ATA will likely replace its parallel predecessor as the standard interface not only for hard disks but optical drives as well.

These are all well-known and highly publicized architectural shifts, but one that may not get the same amount of press is likely to have an even more pervasive impact on system design. Get ready for the serial bus design known as PCI Express.

Like Speed, Only on a Faster Bus

As you might guess from its name, PCI Express (sometimes still called by its codename 3GIO, for third-generation input/output) is a successor to the Peripheral Component Interconnect (PCI) technology seen in desktop machines’ motherboard expansion slots since 1992. (PCI, in turn, replaced older, slower interfaces like ISA — Industry Standard Architecture — and VL-Bus or VESA Local Bus.)

PCI Express is designed to replace not only the PCI bus for devices such as modem and network cards, but the Advanced Graphics Port (AGP) bus used for desktop graphics cards since 1997. Unlike its parallel predecessor, PCI Express is a serial point-to-point interconnect capable of extremely high-bandwidth transfers, ranging from the 250MB/sec of a bare-bones 1X implementation up to 4GB/sec for the 16X PCI Express link likely to dominate the AGP-replacement market.

The 1X and 16X notations refer to the bus width or lanes available. The PCI Express interconnect is also bidirectional, yielding theoretical concurrent bandwidth of up to 8GB/sec for 16X or a whopping 16GB/sec for the specification’s current 32X ceiling.

PCI Express also supports cutting-edge features such as power management, hot-swappable or -pluggable devices, and the flexibility to handle both host-directed and peer-to-peer data transfers. The last is important because it allows PCI Express to emulate a network environment, sending data between two compatible points, without routing it back and forth through the host chip.

PCI Express also simplifies board design, as it’s a single-wire, serial technology with none of the strict wire count or length requirements of parallel bus architectures such as PCI. Scalability is another key feature, as higher-bandwidth PCI Express revisions are planned to far outpace anything PCI or even the server segment’s PCI-X can offer. (More on PCI-X in a moment.)

Since PCI Express also doubles as a chip-to-chip link at the board level, it could theoretically be used to replace a wide range of current bus and interconnect technologies, but is being targeted at fairly specific tasks. For example, despite lots of trash talk in the years before the technologies shipped, we’re unlikely to see a clash of titans between PCI Express and HyperTransport; nor is Intel likely to abandon its proprietary CPU and memory-bus technologies, even as it moves to PCI Express for peripheral connectivity and as a Northbridge-to-Southbridge chipset link.

Standardization in Return for Obsolescence

Besides enabling higher bandwidth, the move to PCI Express will promote bus and connector consistency between components and platforms — not just between today’s PCI peripherals and AGP graphics cards, but between desktop, server, mobile, and other computer species.

Although PCI Express’ board-level implementation costs will be at or below those of PCI, this standardization does come at a price: Though software-compatible with PCI protocols, PCI Express’ physical interface is incompatible with existing PCI and AGP hardware. This poses some challenges for manufacturers and vendors making the changeover, as all the hardware and software must be in place at the right time — without too much pre-launch publicity, lest retailers be stuck with shelves full of soon-to-be-obsolete, hard-to-sell AGP and PCI gear.

Microsoft holds a key piece in this puzzle: operating system support. According to Intel’s and its industry-consortium partners’ white papers, PCI Express is software-compatible with current operating systems, so system initialization, hardware discovery, and resource allocation should be no sweat.

But Microsoft isn’t touting PCI Express support as a major feature of anything short of its next-generation Windows “Longhorn” — and while Longhorn and PCI Express may be made for each other, the former won’t ship until at least 2006 while vendors like ATI and Nvidia are demonstrating PCI Express parts at trade shows and champing at the bit to start selling product. It’s doubtful at this point that current operating systems will be 100 percent compatible with all PCI Express implementations, and it remains to be seen what level of Windows support this year’s first wave of PCI Express graphics cards will have.

A Graphics Windfall

Regardless of the eventual timing and OS support issues, some of the main beneficiaries of the PCI Express revolution will be the abovementioned ATI and Nvidia, as well as other graphics card manufacturers. Since the PCI Express connector will be incompatible with current AGP cards, get ready to open your wallets for a new PC with new-generation graphics — great news for the industry’s bottom line, just as it was when we moved from PCI to AGP video and hot new 3D cards were coming out of the woodwork.

Of course, graphics vendors appreciate the technical as well as fiscal benefits of PCI Express architecture, including higher bandwidth and additional power available to cards. Today’s AGP 8X (a.k.a. AGP 3.0) has really been pushing the envelope in terms of high-performance video cards, and it’s hard to deny it’s time for a change.

PCI Express will alleviate many of the current AGP timing issues and almost triple the maximum power draw, giving the new bus the edge over both AGP and the more esoteric AGP Pro. As the climb from AGP 1X to 8X has shown, additional bus bandwidth has never been the performance panacea the industry promised it would be, but a cleaner signal rate and additional power could yield significant improvements, especially at the top end. A nice side benefit is that PCI Express would make it easy to have multiple high-end graphics cards in the same computer, really opening the door to both hardware and software innovation, although Nvidia, ATI, and friends may need to work overtime on their already Hoover WindTunnel-class cooling solutions.

Looking Back: PCI-X

We mentioned PCI-X, an intermediate step between PCI and PCI Express that’s proven to be a nonissue for desktop users, though the PCI consortium insists it’s not going away as AGP is. PCI-X uses the same basic parallel interface as PCI, while offering higher bandwidth than even many PCI Express implementations (up to 4.3GB/sec for the latest PCI-X 533 flavor).

PCI-X doesn’t scale as well as PCI Express, and brings physical limitations and cost factors to consider as well, but has one advantage — it remains physically compatible with current PCI cards, while PCI Express requires entirely new hardware. For this reason, PCI-X has found a foothold in the server market, since IT managers rightly view servers as longer-term and more cost-intensive propositions than the “use it and lose it” mentality seen in the lower-cost, more commoditized desktop segment.

Looking Ahead: Portable PCI Express and Likely Upgrade Plans

PCI Express won’t be making an impact on desktops alone, but on notebook PCs as well. The venerable PCMCIA and CardBus standards will be replaced by ExpressCard (formerly known as NewCard), which combines PCI Express and USB 2.0 to enable a higher-bandwidth, modular expansion architecture. The ExpressCard specification also supports hot-swappable devices, using either a full-sized 54mm module (accommodating CompactFlash and Microdrive storage) or a slimmer 34mm version.

As with other I/O interconnect shifts, PCI Express won’t take over the desktop (or laptop) overnight. The exact scenario is tough to peg at this time, but high-end PCI Express graphics cards will likely be the first products out the door. This is a natural fit with today’s high-performance 3D cards, not to mention the upgrade mentality of enthusiast buyers and game maniacs, who’ll justify buying a new PC as a way to get the newest CPU socket free with the newest graphics (or vice versa).

For the rest of us, the actual PCI replacement will probably be a bit slower, with a PCI Express Bridge in place to support current PCI-based cards — an interim solution while vendors move today’s onboard (Gigabit Ethernet) and external (removable storage, docking stations) hardware to the PCI Express bus.

The physical PCI Express connector also adds a new wrinkle, as it can take either the form of a familiar card slot or a cable. The cable-connector idea has been gaining momentum over the past year with USB 2.0 and FireWire devices, and its inherent flexibility (think total, small-form-factor PC case redesign) has definite advantages. Still, you can expect the transition to come in waves, and current PCI hardware to follow the ISA precedent and hang on for some years to come.

Categories: Technology