By Network World staff
Network World, 5/18/98

In 1973, who knew? Back then, Bob Metcalfe was just trying to connect a bunch of desktop computers to a single printer. But his work turned into Ethernet, the technology that made modern campus networking practical. In honor of its 25th anniversary, we recently talked to Metcalfe and several other networking veterans about where they see Ethernet heading in the future. They are:

• Bob Metcalfe invented Ethernet in 1973 and founded 3Com Corp. in 1979. He is currently vice president, technology, for International Data Group, the parent company of Network World, and a weekly columnist for NW's sister publication, InfoWorld.
• Vinton Cerf, senior vice president of Internet architecture and engineering at MCI Communications Corp.
• Bernard Daines knows a little something about Ethernet. In 1980, he was a consultant to the team at 3Com Corp. that produced the first commercial Ethernet products. In 1992, Daines co-founded Grand Junction Networks, a Fast Ethernet switch manufacturer that was sold to Cisco Systems, Inc. in 1995 for approximately $350 million. That same year, he founded Packet Engines, Inc., the Gigabit Ethernet switch maker. He is currently the company's chairman of the board, president and CEO.
• Vinod Bhardwaj was co-founder of Kalpana, Inc., the company credited with producing the first Ethernet switch - a switch Bhardwaj designed. He is now president and CEO of high-speed network startup ControlNet, Inc.
• Eric Cooper is CEO of FORE Systems, Inc., the premier vendor of ATM switches. We thought it'd be interesting to get his perspective on Ethernet because he spends so much of his time competing against Ethernet vendors.

Metcalfe:
We're at Gigabit Ethernet now, right? There's clearly work to follow that - it's already going on in places. I suppose we'll eventually get to Terabit Ethernet - that's an easy prediction.

Ethernet used to mean CSMA/CD LAN. Now it means something different. Ethernet is sort of a name for a cadre of companies that for almost 20 years has learned how to build a series of businesses based on standards. So that's what Etherent means - it means that cadre of companies. It's Ethernet in a much stronger, new sense. And it's one of the reasons why Gigabit Ethernet has been kicking ATM's butt recently - because the ATM people don't know how to operate this way. We saw it with ISDN, which was sort of a poor implementation, weakly standardized, slowly deployed, if at all, and then overcharged for, without any competition.

And ATM sort of comes out of that [same group], which doesn't really know how to make standards, it doesn't know how to migrate - it doesn't know how to compete, that's probably its greatest weakness. That would be a systemic explanation for why Gigabit Ethernet has been kicking ATM's butt.

Cerf:
The basic design can work beyond gigabits, but the physical size of the network has to shrink to make it so. Beyond tens of gigabits, it is hard to tell whether it will still be useful, except as a very local interconnect method. Of course switched Ethernet may well work fine at higher speeds since only one device occupies each ``spoke,'' and the switching does what the shared Ethernet did at lower speeds.

Bradner:
Ethernet will continue to get faster - at least on fiber. Early discussions have started on a 10G bit/sec Ethernet standard, but I do not expect to see 10G bit/sec on copper twisted pair anytime soon.

Cooper:
It's already a stretch to call Ethernet a ``technology.'' Now it's mainly just a logical specification of address and frame formats. The early technology - of shared media access, collision detection and avoidance and so on - has been made vestigial, if not quite completely irrelevant, with the advent of switch-based, dedicated-access networks.

As the historical baggage of shared access is left behind, there is no technological reason why Ethernet-compatible address and frame formats can't be used on higher and higher speed dedicated links. It's simply gated by improvements in transceiver technology - electrical, optical or what have you.

Daines:
No, not really I don't. People are already starting to talk about 10G bit. The real beautiful thing about Ethernet over the years has been whenever something new has come along, Ethernet's been able to come along to match it.

Bhardwaj:
There is no limit on how far the technology could go. For fiber, we should see 10G bit/sec Ethernet soon. And for copper, although the current limit seems to be at 1G bit/sec on four pairs. With improvements in cable and by sheer ingenuity of future creative engineers, it will be possible to exceed the current 250M bit/sec per pair limit. Even higher speeds could be obtained by using more than four pairs in a cable. You could have a flat cable with 24 pairs running at 10G bit/sec.

What do you see as the key challenges facing the technology going forward?

Metcalfe:
Getting into the home. In order for ethernet to get to residential in any big way, just as we saw with business, a LAN has to develop first because most of the traffic is LAN traffic and a small fraction leaks out into the WAN. So a big challenge is for Ethernet to evolve on the low end - [to become] cheap and easy to use.

Cerf:
Dealing with higher speeds; possibly rearchitecting to work in optical mode (with optical switches and transverse filters).

Bradner:
Other undefined high-speed LANs.

Cooper:
The Ethernet-compatible frame size of 1500 bytes is a brick-wall impediment to high performance. This is already evident in the Gigabit Ethernet arena, where vendors must resort to proprietary - meaning not ethernet - larger frames to achieve respectable performance.

Ironically, ATM surmounted this challenge years ago by logically separating the notion of small fixed-size cells from larger variable-size frames and using inexpensive silicon to do the conversion. As a result, FORE's 622M bit/sec ATM adapter today delivers more than a gigabit per second of network throughput, in a completely standards-based way, while the fastest of the supposedly Gigabit Ethernet adapters actually deliver less than half that.

Fundamentally, the ethernet standards will have to encompass larger-than-1500 byte packets in order to reach new speeds. But that will limit compatibility with older generations of ethernet. Eventually, ethernet will have become little more than a marketing term, like serial interface.

Daines:
That we don't trivialize it. Technology has become so pervasive and things seem so easy that nobody questions anything. It'llSo, I guess the answer is, just don't assume it'll happen by itself; it'll take work to keep it expanding and growing.

I think in some ways the reason for Ethernet being so successful over the last few years is there have been challenges. 100M bit VG was challenging, so people like [Grand Junction] who were doing 100M bit Ethernet had to work hard to make it happen and do a good job.

Bhardwaj:
For the copper version, it is the limitations of the analog circuitry. A special semiconductor process might be needed just for analog, which could provide higher speed transistors at higher voltages and with better substrate noise isolation. I doubt if the current gigabit standard would work reliably on the installed base of CAT 5 cable and connectors with current state-of-the-art in analog technology.

Will Ethernet have a 50th birthday, or will it be just a memory by the year 2023?

Metcalfe:
I predicted that 2003 would be the last year in which a new Ethernet product would be announced. But I made that prediction based on the 10M bit/sec limit. That was before I saw, as others saw, that you could speed it up. That's been going on steadily. And then when you go switched and eliminate collisions entirely, the sky's the limit. I've changed my mind. And I hope to attend the party. Wouldn't that be nice?

Cerf:
Hard to see that far ahead - it's lasted 25 years and gone from 3M bit/sec to 1000M bit/sec in that time. If it can reach terahertz through switching methods, it could reach 1,000,000M bit/sec by 2023, maybe sooner.

Bradner:
Just like the person who said: ``I do not know what the major programming language will be in the year 2000, but I do know it will be called COBOL,'' I fully expect that a technology called Ethernet will be a major player in networks of the year 2023. But I do not expect that it will be a CSMA/CD-based technology.

Cooper:
Due to its ubiquity, ethernet will survive as a marketing term and perhaps as an emulated compatibility mode supported by the terabit-per-second, combined cell-and-frame, all-switched networks of the year 2023.

Daines:
Unless every computer uses teleportation or has ESP built into them by then, yes, I think Ethernet will have a 50th birthday. Because, after all, what is Ethernet today? It's not black cable looping around computers. It's not even a hub anymore. Ethernet is a well-understood, well-known arrangement of data and some rules that go along with it that is carried using a dozen different media and signaling types, at four or five different speeds. The common element is the frame format and, to some extent, the packet format. No matter how I transmit my signal from one station to another, when it gets back into the frame, you can pass it to another station and [the data] is still there.

Bhardwaj:
It depends on what we mean by Ethernet. Ethernet is today full duplex, and that means it is no longer CSMA/CD technology. It is also no longer a shared medium with the advent of switching. So Ethernet will keep on evolving, but it is quite possible that it is called Ethernet after 50 years even though it will have, in effect, changed from an ape to a human being.

What were your original expectations for Ethernet?

Metcalfe:
The immediate problem was to connect all those new PCs we were building at Parc [Xerox Corp.'s Palo Alto Research Center] to the laser printer we were building. I was given the job of connecting them together. It was the beauty of the mathematics of Ethernet that appealed to me. It's simplicity and elegance at its core. So one of my expectations was to pursue a design based on an elegant mathematical forumation and another one was to hook up those soon-to-be called PCs to that soon-to-be-called laser printer, so that it worked, so I could print out documents at very high speed.

The idea of Ethernet as a standard did not occur until '79, six years later. It was an idea that Gordon Bell at [Digital Equipment Corp.] and I had at a meeting in his office at DEC in February of '79. He wanted me to devise another LAN that DEC could use that wouldn't run afoul of Xerox's patents. And we had the idea, ``Hey, instead of doing all that, why don't we just write a letter to Xerox and propose we work together to make a standard?'' And then it began to look big so that by June, after I brought DEC, Intel and Xerox together, I was convinced there would be an Ethernet-compatible marketplace and that 3Com should be formed to serve it. So it looked big then, in '79, not back in '73 when it was invented - May 22, 1973.

Cerf:
A great and cheap way to interconnect a lot of computers.

What has most surprised you about the evolution of Ethernet?

Metcalfe:
Other than the fact that it passed 100 million connections a long time ago? Other than the fact that 3Com is now a $7 billion company? Those are two big surprises.

I'm surprised about how ugly the process is of standardizing and competing around a standard. It's intrinsically ugly, and it works beautifully. But you have to have a stomach for street fighting. That was kind of a surprise. Innovation is not some artist's conception of innovation. It's really rough and tumble and that's necessary for it to work, and it does work brilliantly. Evidence - Gigabit Ethernet kicking ATM's butt.

Cerf:
Its ability to adapt to higher speeds (through) the switched versions.

Bradner:
The level of assumption by others (first token-ring proponents, then ATM proponents) that Ethernet was a has-been inferior technology and would soon fade to nothing.

Daines:
The biggest surprise I had was at Grand Junction when we were working on a [100M bit/sec Ethernet network interface card] that would start selling at about $1,500, which is about what the FDDI ones were. And I made the real bold prediction that within a few years, we'd get our NIC down to about $250 to $295. Look how wrong I was. It's now $29 or $39 for a 100M bit NIC. I can hardly believe it.