Intel looks past PCI-X
zdnet.com
By Lisa DiCarlo, PC Week Online
September 18, 1998 5:16 pm ET
PALM SPRINGS, Calif. --- While grudgingly accepting the PCI-X specification put forth by IBM, Hewlett-Packard Co. and Compaq Computer Corp., Intel Corp. (INTC) this week launched its own plans for advanced server I/O.
A likely scenario has PCI-X plugging a performance gap for Intel-based servers until switched fabric-based I/O, which Intel is backing, is commercially viable, perhaps in late 2000.
The PCI-X spec, jointly and secretly developed by rivals IBM, HP and Compaq, promises up to a sixfold increase in overall server performance. A PCI-X bus, for example, will run at 133MHz and move up to 1GB of data per second between the CPU and peripherals, sources said.
At the Santa Clara, Calif., company's Developer Forum here this week, Intel executives said they are still evaluating PCI-X but that it seems too focused on bandwidth, not scalability and availability. There is a lot of headroom left in today's PCI, an Intel official said, citing a twofold performance improvement expected in 1999, when Intel introduces a 64-bit-wide 66MHz bus and multisegmented PCI support in the 450NX chip set.
"I haven't been able to understand why they're focused on raw bandwidth," Justin Rattner, an Intel fellow and director of Intel's Server Architecture Lab, in Hillsboro, Ore., said of PCI-X. "Where is the improvement if all you do is wait faster for your data?"
Despite the criticism, sources said Intel has already given a nod to the PCI-X trio.
An executive with one of the companies involved in developing the spec said they designed PCI-X to address PCI's current power and performance limitations. He agreed that a fabric-based architecture is the way to go long-term. "After about the year 2000, we will need a very different architecture to handle I/O," the executive said.
The PCI Special Interest Group, a standards body, is evaluating PCI-X and will make a determination in October on whether to support it.
Switched fabrics are vastly different from buses. Rather than take a signal from a CPU into a bus through hard wires and into an I/O subsystem (the PCI bus implementation), signals come from the chip and into a high-speed switch. This switch acts as a controller and directs the signal to a subsystem, another chip or even another server. Performance is based on the speed of the switch.
There is far less latency in switched fabrics because CPUs do not waste power pushing data through a PCI pipe.
"We will move away from bus-based solutions to switched fabric," said Intel's Rattner.
Intel is working with a very small group of companies on fabrics, possibly as few as one. It will host a large-scale private meeting in November for companies seriously interested in supporting fabrics.
"The next wave [in I/O] will be switched fabrics. They should replace buses," said Jerry Sheridan, an analyst at Dataquest Inc., in San Jose, Calif.
Supercomputers and mainframes have employed custom switched fabric implementations for years, but Intel's plan is to make fabrics extensible across all Intel server hardware at a volume-based cost.
The challenge will be in convincing server makers that its plans are not exclusionary or proprietary, which is what drove the development of PCI-X in the first place. |
