July 21, 2003
(Computerworld)
Users comparing throughput numbers for PCI Express and PCI-X may find the experience confusing. That's because the nature of the two architectures makes apples-to-apples comparisons more difficult.
Since Express is a serial technology, designers tend to describe bandwidth in bits per second. In contrast, PCI designers describe bandwidth in bytes per second. And even after calculating the bit/byte conversions, the bandwidth numbers aren't directly comparable because the ways in which the two technologies transport data are different.
A PCI Express design transmits data over a set of two wires at a base signaling rate of 2.5Gbit/sec.—but only in one direction. A complete, full-duplex PCI Express "lane" consists of four wires, for an aggregate bandwidth of 5Gbit/sec. Although PCI-X 2.0 adds bandwidth by increasing the clock speed to double data rate (DDR) or quad data rate (QDR), system designers increase the bandwidth of a PCI Express slot by adding lanes. The range starts with a 2.5/5Gbit/sec., one-lane or 1X (pronounced "times one") connection all the way up to an 8/16Gbit/sec. 32X design. Initially, desktop expansion slots are expected to range from 1X to 4X; servers will likely range from 4X to 8X; and graphics will run off 16X slots.
So while PCI-X 2.0 supports up to 2GB/sec. of traffic in either direction, a 4X Express slot supports 2GB, but only 1GB/sec. in one direction. If the nature of the I/O is unidirectional—say read-only access to a serial-attached storage array—then a comparable PCI Express device would need to be 8X.
Note: All bandwidth numbers shown in Gbit/sec. for comparison purposes.
Server-based I/O technologies are leading the charge in higher I/O bandwidth requirements. Bottlenecks could be compounded for storage and networking technologies, because more than one adapter may be used. On the desktop, graphics technologies with bandwidth requirements beyond the capacity of AGP 8X will use PCI Express.
Source: Serverworks Inc.