I had problems linking to that IXP article, but I found it eventually. In case anyone else has the same trouble, here is the text:
Intel makes IXP its net processor cornerstone
By Craig Matsumoto
EE Times
(08/25/00, 2:26 p.m. EST)
SAN JOSE, Calif. — One year after unveiling its network processor and communications focus, Intel Corp. is claiming success on both fronts and is moving to expand its networking operations.
The company's new attitude toward networking is particularly evident with the network processor. A Digital Semiconductor project that was nearly killed when Intel acquired that rival's chip business in 1998, the IXP1200 has become a key part of Intel's strategy and will be expanded to cover every stage of the network, from simple access devices to high-speed core routers. Intel also is readying an applications programming interface (API) for the processors to aid in software development.
That the IXP1200 is alive at all is due largely to the efforts of Mark Christensen, vice president of Intel's network communications group. Through a combination of foresight and luck, Christensen got Intel to back the part, which has become the star product in the Intel portfolio of networking silicon.
At a small gathering with press and analysts this week, Christensen took time to reflect on Intel's new networking face. Evidence of the shift could be seen in the attendance roster for this week's Intel Developer Forum, for which 1,500 networking engineers were on hand. "A year ago I got, like, 30," Christensen said.
Linley Gwennap, principal analyst with The Linley Group (Mountain View, Calif.), said Intel has surpassed his expectations in its ability to produce not only the chip but the requisite tools and software as well. "When [the IXP1200] first came out it looked like a good technology, but I thought it was going to be tough for Intel to put together all the pieces to get into this market. I was pretty skeptical. But they've done a good job pulling it together," he said.
Still, Intel's success in networking isn't etched in stone just yet. Robin Melnick, director of marketing for competing network-processor vendor MMC Networks Inc., noted that the bulk of Intel's design wins have been in network appliances — which in many cases originated in the PC environment, Intel's stomping ground.
"Intel is the 800-pound gorilla of PC processors. Their approach to networking seems to be to take the halo effect from that," Melnick said. "In networking, it's a completely different game."
As Intel moves into high-end networking, analyst Gwennap said, its brand name will hold less sway, and its products will face demands — such as high availability — that are unknown in the PC world. "Intel has a better understanding of the low end," he said. "Companies like IBM have been selling processors into high-end [networking] servers for a long time."
Christensen has held his post since 1990 and is part of a triumvirate of executives leading Intel's new direction. A reorganization six months ago fit their responsibilities into neat categories: Christensen handles chips, John Miner works with networking systems and Ron Smith is responsible for wireless technologies.
Intel claims 60 design wins for the IXP1200, a figure that Christensen guesses is higher than design wins for all other network processors combined. Its design group has grown from what Christensen called "a handful of people" to encompass a next-generation team in Santa Clara, Calif.; another team in Chandler, Ariz.; a group in Fremont, Calif., targeting low-end designs; and three parallel teams in Boston.
The IXP1200 design group began life as a project of Digital Equipment in Boston and wound up in Christensen's hands after Intel's acquisition of Digital's semiconductor operations. At the time, the idea of a network processor hadn't hit the market; packet forwarding was still handled by ASICs.
Death sentence reversed
Christensen was advised to shut down the Digital project, because the part was too complex and too far from profitability. "I had all my guys telling me to kill this thing," he said. But one customer talked about the processor as if it were a miracle, telling Christensen it was exactly what networking OEMs were going to need.
Christensen pleaded his case before Intel chief executive Craig Barrett. Christensen acknowledged that the excursion would cost tens of millions and require at least a five-year commitment, but he dangled the potential reward of prominence in the lucrative networking market. Barrett, who was in the process of shifting Intel's sights from PCs to communications, green-lighted the idea.
Christensen was lucky that his network processor team chose to stay after oversight of the team shifted from Digital to Intel. Digital had always treated Intel as its "Darth Vader," Christensen said, and the culture clash made the acquisition difficult. Many of the Digital faithful, including the team working on Digital's StrongARM processor, had walked out.
Christensen pursued a mainstream market for the processor, targeting the OC-3 (155-Mbit/second) and OC-12 (622-Mbit/s) Sonet grades rather than the more cutting-edge speeds. "We decided not to go after the optical network first, like some of our competitors did. That's not how you establish an architecture," he said.
Possibly as a result of its less ambitious targets, Intel has joined MMC as one of the few network processor companies shipping product in volume. That means Intel already has first-generation field test data in hand. "We have a great idea what we did wrong," Christensen said. "The other guys have no idea."
MMC, however, interprets Intel's moves differently. "They're conceding right now that they have no play in the market we see as the sweet spot," namely edge routers and optical networking, Melnick said. "They've got one chip that's too expensive to play at the low end and too underpowered to play at the high end."
High-end, low-end incarnations
Intel does intend to expand beyond the IXP1200. Plans outlined at the Developer Forum included versions to cover low- and high-end systems, as well as an API slated for release this year.
Frank Schapfel, manager of Intel's network processor operations, said the 200-MHz version of the IXP1200 is shipping in volume now and will be followed later in the year by an enhanced IXP chip, details of which were not specified.
The IXP1200's StrongARM processor will be replaced by the XScale, the StrongARM successor developed by an all-Intel team. The XScale's ability to run at low power — reportedly 1.5 W for 1-GHz performance — is going to be a critical advantage for Intel, Christensen said.
"I believe the ultimate sweet spot is a 5-W package, because in a 5-W package you can go into everything from a CPE [customer-premises equipment] box to a gigabit blade," he said.
Second- and third-generation processors based on XScale are due next year, Schapfel said. Also due in 2001 is a high-performance version of the IXP, targeting the high-speed network core.
With those additions, the IXP family will cover equipment at almost every stage of the network, except the customer premises. Intel is relying on its acquisition of Basis Communications Corp. to bring its networking products — collectively called the Internet Exchange Architecture (IXA) — into that market.
Basis will develop "service-specific platform" chips, the first of which, the BCG6911, will target digital subscriber line equipment. The chip will sit between the DSL physical-layer parts and the Ethernet connection that leads to the carrier network.
Intel is betting its wide-ranging approach will carry the architecture to multiple contracts within large companies such as Cisco, Nortel, Lucent and Alcatel, Christensen said.
Customers like to see a wider-ranging road map, analyst Gwennap said. "That's probably why Intel has been able to get so many design wins, because people see the inherent scalability of the architecture."
Intel also is putting emphasis on software, providing not only programming tools but basic libraries to cover fundamental routing functions. Attention to software is vital to survival in this business, Christensen said.
Hard software lessons
Melnick agreed, noting that MMC had learned about software the hard way. The company initially paid little attention to software, expecting to make sales on the merits of its chips. That attitude changed as it became clear that MMC's software had to be at least on a par with the competition's. "This year, more and more, we're looking at software being a compelling advantage," Melnick said.
Like most network-processor vendors, Intel is developing APIs to allow OEMs to reuse software. "The goal there is to have the software companies and the equipment companies write to the APIs so that as we change underneath, they don't have to rewrite their software," Christensen said.
IX-API being readied for release later this year will include Intel's IX symbolic language for writing IXP microengine code, OS-independent run-time software, libraries and tools.
To assist in programming the network processors, the API includes software modules called action classification engines (ACEs), which can be written by Intel or by outside parties. The engines come in two varieties: "conventional" ACEs, intended to work with any IXA hardware, and "micro" ACEs, targeted at specific microengines.
Each ACE is a prepackaged programming task akin to a subroutine. Ideally, ACEs will be written to handle common tasks, such as packet parsing or network address translation, and the necessary ACEs will be strung together to form a larger software routine.
Details on availability schedules for the ACEs will be released next month at Networld+Interop, said Larry Huston, an Intel software architect. Intel has pledged to include some of its own ACEs, covering extremely common functions, in the first release of IX-API.
To ease development of micro-ACEs, Intel will offer macro libraries with receiving and transmitting code to let a micro-ACE connect to other ACEs, essentially forming a prefab wrapper for the developer's code. |
