Single page Print

PCI Express for graphics

Analyzing ATI and NVIDIA's PCI-E strategies
— 12:00 AM on March 26, 2004

ALTHOUGH motherboards and core logic chips with PCI Express support aren't due out for weeks, ATI and NVIDIA have already announced plans to make current generation Radeon and GeForce FX parts compatible with the new interconnect technology. ATI and NVIDIA agree that PCI Express is the next big interface for graphics, but the two companies are initially supporting PCI Express in very different ways: ATI will provide PCI Express compatibility with a new line of GPUs that offer native PCI-E support, while NVIDIA's first PCI Express efforts will use a bridge chip to graft AGP GPUs to the PCI-E interface.

According to ATI, bridging creates all sorts of performance and compatibility problems that won't impede its native PCI-E implementation. However, NVIDIA insists that its bridge solution is sound. Join me as I dissect ATI and NVIDIA's PCI Express strategies to see if either argument holds water.

PCI Express basics
Before we dive in, it's worth taking a few moments to go over what exactly PCI Express is and what the technology will bring to the table for graphics.

PCI Express is a third-generation interconnect technology that's slated to replace PCI, PCI-X, and AGP. To avoid performance problems that can arise from bandwidth sharing on a common bus, the PCI Express architecture is based on a series of point-to-point connections. These connections employ a serialized, packed-based data transfer scheme more similar to an Ethernet network than to a traditional PCI bus. Each of these connections has dedicated upstream and downstream bandwidth and supports bidirectional transfers. In cases where more than two devices need to communicate with one another, chipset makers can build a switched fabric to arbitrate multiple PCI-E conversations without compromising bandwidth to any device.

To meet the varying bandwidth needs of different system components, PCI Express can scale from one to 32 "lanes". A single-lane X1 link yields 250MB/sec of dedicated bandwidth in each direction, which should be enough for mainstream audio, networking, and storage applications, but not nearly enough for graphics. To replace AGP, the first wave of PCI Express graphics implementations will use a sixteen-lane X16 link that offers 4GB/sec of bandwidth in each direction—much more than AGP 8X's 2.1GB/s total bandwidth. Although X4 and X8 links are also available, X1 and X16 should be the most popular for personal computers. Expect the first PCI Express-enabled core logic chips to sport a handful of X1 links hanging off the south bridge and a single X16 link tied to the north bridge.

There's more to PCI-E than just additional bandwidth. Thanks to its serialized design, PCI Express delivers much higher bandwidth per pin than legacy I/O standards, which helps to reduce costs and enable smaller form factors. PCI Express also supports advanced power management and hot plugging, and its packet-based protocol enables isochronous transfers for time-dependent data delivery and quality of service arbitration to optimize bandwidth for high-priority data streams.

With all sorts of new features and plenty of bandwidth to spare, PCI Express is a huge jump over both PCI and AGP. However, PCI Express also retains compatibility with PCI software by retaining the PCI device initialization and memory models, which means that drivers and operating systems won't have to undergo major retooling to support the new interconnect.

ATI goes native
ATI's PCI Express strategy relies on a new line of native PCI-E GPUs that will mirror the company's current AGP graphics products. With the exception of differing on-chip interconnect interfaces, these new PCI-E GPUs will be identical to their AGP-based counterparts. ATI claims its native approach to PCI Express support allows its GPUs to take better advantage of PCI-E's performance potential by cutting out the latency inherent to bridged implementations. ATI's PCI-E GPUs will deliver PCI Express X16's 4GB/sec of upstream and downstream bandwidth in full, something ATI suggests bridged implementations may not do.

ATI also suggests that its native approach is more economical than bridging because there are no additional bridge chip costs. However, ATI makes no mention of possible costs associated with maintaining parallel AGP and PCI-E GPU production.

ATI's strategy should make board design easier and more economical for ATI's add-in board partners. Card makers won't need to worry about the additional cost and complexity associated with implementing a bridge chip, and they won't have to worry about the bridge as an extra potential point of failure. For manufacturers contemplating whether to partner with ATI or NVIDIA, the simplicity of ATI's native PCI-E implementation will surely be attractive. However, with most graphics card manufacturers building or rebadging vanilla reference designs already, we're unsure how much of a factor such considerations will be.

ATI also claims that bridged PCI Express solutions will have a slight power-on delay when compared with native implementations, and that bridged solutions may not support all of PCI Express's features, most notably advanced power management. ATI made this claim about advanced power management before NVIDIA announced the details of its PCI Express bridge, and as it turns out, that concern may have been a bit premature.