DSP + PCI-X = Processing Bandwidth

The time has come. DSP board and system vendors have a decision to make: to move to the PCI-X bus or not. PCI-X, the extension of the venerable PCI bus, will happen. The infrastructure is falling in place to support PCI-X and the system makers are on their way to incorporating PCI-X into their servers. For them, PCI-X offers higher and more stable bandwidth, overcoming some of the key weaknesses of the original PCI bus.

That's the good news. The bad news is that PCI-X is not 100% compatible with the existing PCI bus. Yes, it does run existing PCI cards or boards. But in order to do so it degrades its performance so that it's just another PCI bus variant. The result? No performance advantage. So it's really PCI-X or PCI, take your pick. But you can't have PCI compatibility (i.e., running PCI cards/boards) and PCI-X performance.

For awhile it looked like PCI-X wasn't going to really happen. For one thing, it was developed by the Big Three—Compaq, HP, and IBM and their allies, not Intel. Intel, who basically single-handedly developed PCI technology, was not happy about PCI-X, especially when PCI-X was caught up in the Intel's NGIO Vs Big Three's FutureI/O wars.

Intel definitely has a NIH attitude when it comes to PCI and its architecture initiatives. And PCI-X was not on its long-term Pentium chipset roadmap. The problem is that Intel is crucial to the acceptance of a PC hardware technology for Intel carries a lot of weight as the leading PC uP vendor, leading PC chipset vendor, and leading PC motherboard vendor.

Also, PCI-X was perceived by many to be a system company's solution to its server I/O bottlenecks, a bus narrowly defined by system companies like Compaq, HP, and IBM. It was not expected to be a robust, multi-level bus like PCI with heavy application in the embedded space. For while the PCI bus has a lot of drawbacks, it has become the ubiquitous bus, with a PCI variant in every 32-bit or bigger system.

PCI-X is now on a roll. Perhaps folks figured it was better to fix the PCI bus, rather than keep trying to go faster and wider. Infrastructure is now falling in place to make PCI-X a reality, not just for PC server systems, but also for embedded systems as well. PCI-X controller cores and chips are on the way, including FPGA implementations.

PCI-X looks to be a winner, and not just as a PC server I/O bus. Work is underway to explore a CompactPCI version for embedded system servers and platforms. Current research findings show that a PCI-X running at 66 MHz and 64-bits wide can drive between five to seven slots (relaxed timing), opening it up for cPCI class operation.

In engineering, there is old joke: "You know when a technology is going mainstream; that's when the tester people decide to build equipment for it." Well, the bus tester folks are now climbing on the PCI-X bandwagon, providing test equipment for PCI-X implementations. Agilent, Catalyst, VMEtro, and others are all anteing up PCI-X test equipment. The wind is blowing right. When you build PCI-X-based equipment you'll be able to test it.

The DSP community was slow to adapt to PCI. For one thing it was the PC's I/O bus. For another, many board and system vendors had proprietary buses or were using extensions of their DSP's main bus. Nevertheless, most people liked the advantages of using a PC standard product technology, which included low cost, wide product selection and software compatibility. They've learned to live with PCI's flaws and reap its benefits.

Today, another choice is emerging. To PCI-X or not. On the positive side, PCI-X offers a PCI done the way it should have been: split transactions, large status word, byte-count transactions, wait-state restrictions, and so on. On the negative side, PCI-X means leaving the PCI cocoon with a new standard. But it may be worth it. There's a good chance that PCI-X will follow the PCI path and emerge as a universal bus at multiple levels—System (cPCI), I/O (PCI), Mezzanine (PMC, PC-MIP). PCI-X has what DSP engineers need—bandwidth. It's time to choose.