The fabless phenomenon

'Fabless semi' firms are building net functions into silicon, paving the way for high-powered, low-cost network gear.

By Todd Wallack
Network World Fusion, 3/2/98

In a sunny office tucked away in Framingham, Mass., a small group of engineers are quietly designing chips for Cisco Systems, Inc.'s flagship routers. As they work, they sketch out block diagrams with markers, creating a tangle of digits, acronyms and arrows. It looks like a football play where the players are bits of code scrambling for touchdowns in a silicon stadium.

But there is no Cisco sign out front. And the engineers, who work in a start-up just upstairs from a sand and gravel company, are also working on nearly identical chips for 3Com Corp., Bay Networks, Inc. and other network equipment giants.

Welcome to Maker Communications, Inc. and the world of fabless semis. Scores of firms you've never heard of are quietly creating low-cost chips that handle all kinds of network functions, from physical connections to wire-speed routing. Like Maker, most of the firms design semiconductors - or semis in industry parlance - but hire outside foundries to fabricate them, hence the term fabless semis. Most of these firms are relatively small, but the effect they may have on the network industry is anything but.

"These guys are building the next generation of LANs," says Frank Dzubeck, president of Communications Networks Architects in Washington, D.C. Dzubeck believes the firms are gradually turning network chips into a commodity, where price is the only thing that matters. In the process, they are forcing the likes of Cisco and Bay Networks to focus more on service, distribution and advanced software features. "It's moving the battleground upscale," Dzubeck says.

Moreover, by focusing on network technology, the fabless semi companies could spawn a slew of start-ups that can churn out low-cost hubs and switches by packaging specialized chips with off-the-shelf power supplies and casings - the network version of PC clone makers.

"This provides a low-cost alternative that is ideal for companies that don't have the resources to design the chips themselves," says Diane Myers, a senior analyst with In-Stat, a semiconductor market research firm in Scottsdale, Ariz. "They can get to market quickly and offer equipment cheaply."

Fabless players

To be sure, fabless semi companies are nothing new. In fact, there is a well-established trade group, the Fabless Semiconductor Association, that has 160 member companies. But more and more of those companies see an opportunity to build network functions into silicon. The fabless group estimates 64 of its members, or 40%, are involved in the network industry.

Maker is one of these upstarts. Despite the prowess of the big network equipment vendors, which also design their own silicon, 3-year-old Maker was the first to come up with a chip that could convert data between cell and packet formats at speeds up to 622M bit/sec.

In a business where time to market is of the essence, none of the big equipment vendors wanted to risk using a slower chip or take the time to develop equivalent technology on their own. Consequently, you can already find Maker's CellMaker chip in Cisco's 12000 (formerly known as the BFR) and Catalyst 5000 products, among others. Cisco, Bay and 3Com all confirmed they signed deals with Maker in order to quickly get the cell-to-packet conversion feature in their products.

"In some ways they are pushing the envelope," says Bob Lapointe, Bay's senior manager of value engineering. He confirmed Bay has a relationship with Maker but declined to identify specific products involved.

Maker founder Bill Giudicesays his company's early sales - $5 million to $10 million last year - are just the beginning. Equipment vendors are just starting to buy network chips from outside designers, and he expects the market to boom. Maker even has its sights set on an initial public offering (IPO) sometime in 1999 or so.

Another fabless semi company that has grabbed analysts' attention is MMC Networks, Inc., in Sunnyvale, Calif. Though not the largest of the fabless companies, MMC claims it has designed 35 products for more than 27 network equipment vendors. Its ATMS2000 chip, for instance, is used in Cisco's low-end ATM switches.

The Cisco deal in particular helped give MMC enough momentum to launch an IPO in October, issuing 4 million shares for $11 each, raising more than $40 million after fees. The company grossed about $10.5 million in 1996 and more than doubled that mark in 1997, with $21.9 million in revenue. Profits nearly doubled last year to $1.2 million. As of September, the 5-year-old company had 84 employees, more than half in research and development.

Cisco accounted for half of MMC's revenues and Hitachi for another 10%. But MMC is trying to broaden its reach, launching an array of products it claims will form the core silicon engines of LAN and WAN switches handling ATM and Fast Ethernet.

"We give the customer the basic building blocks, and they can offer their own features on top," says MMC chief executive Prabhat K. Dubey. Those building blocks include the following:
The PS1000 Fast Ethernet network processors. MMC says these chips can be used for wire-speed routing. The chips support from eight to 128 10M bit/sec Ethernet switch ports and up to 16 100M bit/sec ports, plus one or two ATM uplinks. MMC has already introduced at least sevenchips in the product family.

The ATMS2000 ATM network processors. These chips are designed for switches that support campus backbones, workgroups or WAN access. They support speeds of 2.5G to 5G bit/sec with 32 OC-3 ports or eight OC-12 ports. MMC has launched at least seven types of chips in this group.
AnyFlow 5000 network processor. This chip supports ATM and Fast Ethernet switches. It features Layer 3 switching, quality of service (QoS) functions and packet/cell interworking at 20G bit/sec with a throughput of 20 million packets per second. Samples of the chip are just starting to become available.

MMC also touts its patented ViX architecture, which features point-to-point connections and centralized shared memory. The architecture is designed to support per-flow queuing technology to buffer information during periods of congestion, direct replication engine technology for multicast signals, virtual segmentation and reassembly to convert data between frames and cells, and programmable bitstream processors to make it easier for network equipment vendors to program additional features into MMC's chips..

If all that sounds familiar, it's because those are the some of the same functions switch vendors have been furiously trying to build into their equipment. Now, fabless semi companies are making them available in generic chips.

As more of these features become standard, analysts say, equipment vendors can either reduce their own development costs or shift their resources to even more advancements.

Another fabless semi firm, Integrated Telecom Technology, Inc. (IgT) in Gaithersburg, Md., claims a large network equipment vendor shut down development on its own chip after seeing IgT's segmentation and reassembly chip in action.

"They were flabbergasted," says Greg Werth, vice president of marketing, who declined to name the equipment vendor. "The only way to succeed is to show them significant intellectual property."

Pricing picture

Whether this wealth of off-the-shelf technology will translate into lower prices remains to be seen, but the potential certainly exists. For one thing, as Maker's Giudice points out, fabless semi firms can spread out development costs by selling to multiple customers. It costs his company $1 million to $2.5 million to design a chip - costs that a company such as Cisco or Bay would be forced to eat if it developed its own chip.

Tim Lindenfelser, vice president of marketing at Broadcom Corp., a 7-year-old fabless semi firm in Irvine, Calif., says his company has been focusing of late on Fast Ethernet chips. In a 12-port Fast Ethernet hub configuration, for instance, the firm was able to combine functions normally handled by 25 chips onto four. Specifically, Broadcom modified the physical interface chips so that each can handle more ports.

For network equipment vendors, that reduces the price from $10 per port to about $5. And the company can knock another $1 per port off the cost by consolidating the repeater function onto the remaining physical layer chips. Multiply that $6 per port savings by 12 ports and you get a reduction of $72 per hub at the manufacturer level.

Lindenfelser estimates each $1 saved in parts translates to a $5 reduction in the final cost of a product, although Dataquest, Inc. analyst Greg Sheppard says $3 to $4 is probably more accurate. Depending on which number you use, that translates into a $216 to $360 reduction off last year's $1,320 retail price for a 12-port Fast Ethernet hub. "It's a huge cost difference," Lindenfelser says. The chips are already available and going into 3Com and Bay products.

Going forward, Broadcom figures it can squeeze even more dramatic savings out of switches. In addition to the $6 savings per port ($18-$30 retail) on the physical chip level, Broadcom predicts it will be able to consolidate media access control functions by the end of this year, reducing the manufacturing cost by another $8 ($24 to $40 retail). That $14 savings at the manufacturing level translates to a total retail cost reduction between $42 and $70 per port. That would be a 15% to 25% reduction in the overall cost of a Fast Ethernet switch, based on the 1997 rate of $275 per port, Lindenfelser says. "The analysts don't quite understand the price reductions ahead," he notes.

Sheppard notes chip prices have already been plummeting. Fast Ethernet chip prices are down 35% to 40% from a year ago, in part because of new chip makers. He projects prices to decline another 20% this year. "NICs and switches have been dramatically impacted by chip price moves," he says.

Justin Smith, an analyst with International Data Corp., also sees fabless companies pushing down prices at the low end, such as Ethernet and Fast Ethernet network interface cards and hubs. "Everything from the client up to the wiring closet is definitely becoming a game of who can produce the equipment least expensively," Smith says. "The chip companies bring a lot to bear on that equation."

Smith expects Ethernet card prices, for instance, to fall by more than one-third by the year 2001. Even today, he says unmanaged 10M bit/sec Ethernet hubs can be had for less than $10 per port. In many cases, they are essentially given away in deals for other equipment.

At the higher end of the market, start-up Orologic, Inc. is also trying to consolidate features onto fewer chips. Raif Onvural, co-founder and vice president of engineering, predicts network equipment vendors will be able to build a 16-port OC-3 ATM switch for less than $85 per port. He estimates the retail price would be $170, though the price would actually range from $255 to $340 per port using Sheppard's multiple. In any case, either estimate handily beats the $400 to $500 market price today, Onvural says. And he says his box will contain more memory and features than any campus backbone switch currently on the market.

Onvural estimates network equipment vendors could shave 20% to 30% off prices by using off-the-shelf chips with QoS and other features instead of developing them on their own. That's based on the assumption that the development of the chip would require 15 engineers or more and cost $2.5 million to $3 million per year, and that companies like Orologic could spread the development costs over several users.

Luke Szymczak, a network industry analyst at Prudential Securities in New York, says the rise of network chip makers is a natural evolution. It doesn't make sense, he says, for the Bays and Ciscos to continue to spend money developing components internally if they can buy them on the open market. Instead, he says, equipment vendors need to focus on pioneering advanced features to stay ahead of the competition.

Building better boxes

To varying degrees, equipment vendors have been taking Szymczak's advice.

Cisco, for instance, says it wants to focus on software and other strengths and welcomes the chance to buy more generic network chips. "It's a good thing if it allows us to get products to our customers faster, cheaper and with better performance," says Marc Beckman, senior manager of semiconductor commodity management at Cisco.

Bay and 3Com, however, while acknowledging that they do use some silicon from fabless semi companies, profess that they plan to continue designing most chips on their own.

"I think there are some features that we are looking to incorporate into our products and, in most cases, we don't see those features being available as off-the-shelf products," Bay's Lapointe says, noting Maker is one of the exceptions.

In some cases, however, equipment vendors like the fabless semi chips so much they buy the company. Cisco, for instance, last June paid $89 million for Skystone Systems Corp., a fabless semi firm in Ottawa that specializes in integrating telecommunications functions for fiber-optic networks. And in December 1996 Bay snapped up NetICs, Inc., a fabless company that designed an innovative Fast Ethernet chip, for $99 million.

For its part, 3Com says it will typically look outside only for the most primitive network components. It likes to embed its own code in the switching fabric and other advanced chips.

"The problem comes as you go up the stack," says Robert Ciampa, 3Com's product technology manager. "These [fabless semi] manufacturers want to give you a total turnkey solution and we have specific software" that conflicts with that.

Ciampa concedes 3Com could load its own software onto a programmable chip, but says that is usually more clumsy than designing the chip internally. He also said equipment vendors are already working on consolidating functions onto fewer chips and taking other steps to slash costs on their own.

The long term

Dzubeck thinks the real effect fabless semi companies will have on the network industry won't be felt until the year 2000. "It's going to take a little while to develop the market," he says, noting that companies such as Bay and 3Com aren't likely to do anything to accelerate the process. "How fast do you want to drive down your prices?" he asks. Eventually, though, he envisions LAN chips selling for "less than the cost of the power supply." And he sees fabless companies tossing in extra features, like Layer 3 switching, practically for free.

Wall Street is already betting several fabless semis are going to cash in. "It's not conjecture," says Clark Westmont, an analyst with NationsBanc Montgomery Securities LLC. The market has already taken off, he says, noting companies such as PMC-Sierra, Inc., which in the last quarter of 1997 saw 91% revenue growth in its network business compared with the same period in 1996. "It's already happened."

Just how much of an impact the companies will ultimately have is hard to gauge.

3Com's Ciampa said he could see start-up switch companies, most likely in Asia, building vanilla boxes to undercut the established players. But he maintains that 3Com and others are cutting prices so much that an upstart would have a hard time making a dent in the market. Plus, companies like 3Com have other advantages, including an established sales and service force, not to mention advanced software functions.

Dzubeck, on the other hand, thinks the effect will be more dramatic. He foresees a world of dirt-cheap boxes that have functions like Layer 3 switching tossed in essentially for free.

Given the pace of change, it may not be all that long before Dzubeck's vision becomes reality.

Maker, for example, has been in business for less than three years, yet Giudice says he has lined up deals with four of the largest data communications equipment companies, meaning vendors such as Cisco, and several large telecommunications equipment makers, including Northern Telecom, Inc.

And Maker has grown from two people working out of their homes - Giudice and co-founder Paul Bergantino - to a cubbyhole in Waltham, Mass., to the current 43-person office in Framingham.

But Giudice and other fabless pioneers don't like to talk about the present. They prefer to dwell on what the industry will be like tomorrow, when you'll be able to buy less expensive equipment with better performance and features. It'll be a world created at least in part by fabless semi companies that now you have heard of, companies like Maker Communications.