Andy's forte is workstation system design, not networking. He designed
the original Sun-I while still at Stanford, the Sun-II, which is the
first Sun workstation most people have seen, and "Dragon" which became
the first Sun SMP.
Unclear what to make of the exchange below. Andy makes a lot of
sense in attacking ATM, but as Rohit would say, you can never trust
my assessment of an article after a single read.
Andy Bechtolscheim, Sun Micro's co-founder, predicts the death of ATM.
In Hollywood, disaster pictures are the rage. On the heels of Twister,
which has already grossed over $100 million, movie executives are
developing motion pictures about tsunamis, volcanos, and forest fires
almost as quickly as they can find them. This kind of feeding frenzy may
sound extreme, but a similar sort of mania is taking hold in the venture
In Silicon Valley, however, it is Gigabit Ethernet--not disaster
movies--that is causing investors to lose sleep.
In the past year, Gigabit Ethernet has captured the interest of venture
community by raising questions about the future of ATM (Asynchronous
Transfer Mode), a switching technology that was especially designed for
mixed media. According to its proponents, this new networking technology
is revolutionary because it promises to build upon the already existing
Ethernet infrastructure. Gigabit Ethernet is so hot that it has already
spawned a string of new startups, including Allegro, Packeteer, Packet
Engines, Rapid City, and an as yet unnamed company headed by Bobby
Johnson of Centillion fame.
One company that has attracted a great deal of attention from VCs is
Granite Systems(so-called when we went to press). This Palo Alto-based
startup has risen above the noise because of co-founder Andreas
Bechtolsheim, Sun's first employee. With co-founder David Cheriton, who
is a professor at Stanford University; Kleiner Perkins, with whom Mr.
Bechtolsheim is an investor; and an unnamed "systems company" (rumored
to be Sun), the tall German emigre is betting $5 million of his own
money that Gigabit Ethernet will soon surpass ATM. Wearing his trademark
Birkenstocks, Mr. Bechtolsheim explained why Gigabit Ethernet will
shortly vanquish ATM.
THE HERRING: How did the company get started?
Bechtolscheim: About a year ago, I called David Cheriton at Stanford and
told him that I couldn't take this ATM bullshit anymore. ATM is a scam
of the phone companies, which are trying to get back into the data
networking business, which they missed out on. And it doesn't work! So
we started talking and very quickly concluded that the most obvious
thing to do to make networks work better was simply to scale the speed
of the existing network technology. So the two of us started a company
to develop products for Gigabit Ethernet. A year ago, Gigabit Ethernet
was a glimmer in our eyes. We weren't sure back then how long it would
take for there to be a standard. But as it turns out, the IEEE picked up
the standards work in March of this year, so the timing could not have
THE HERRING: So you're betting completely against ATM?
Bechtolscheim: Oh, ATM is dead. It's not even a question worth
discussing. What people missed is that what matters is IP, not ATM. IP
is the network protocol that everybody is converging around. And IP is
completely independent of the underlying network technology. You can
run it on Ethernet, Token Ring, FDDI, and even ATM. It's just
particularly difficult to make IP work on ATM.
But if you look at it in reverse, the network technology that will win
is the one that is the simplest, fastest, cheapest, and easiest to use
with IP. And that's Ethernet, not ATM. It has nothing to do with ATM's
supposed technical superiority. The market confusion began in 1993,
when the ATM industry began sending out videotapes proclaiming that
ATM was the future of networking. It was going to be scaled and switched,
with video and this and that, and it referred to IP and Ethernet as the
THE HERRING: Who was sending out these videos?
Bechtolscheim: It was the ATM Forum. Yet five years later, after twelve
technical working group committees and countless meetings and an
unbelievable amount of paperwork, there still isn't a standard that
singly dynamic IP environments.
ATM was supposed to be great for multimedia, but it turns out, it's
basically incompatible with IP Multicast, which is the underlying
technology that people use to deliver multimedia; there is no mapping of
IP Multicast to ATM short of building a replication server that sends
out little packets for each subscriber, which is not multicast but
replication. If you look at the people trying to do video applications
to the desktop like Judy Estrin's company--Precept--they are making
multimedia work over Ethernet because there is no market for video to
the desktop over ATM. Application developers have to use what's out
there, which is Ethernet. And video works just fine over Ethernet. It's
no big deal. It's just people didn't think about it before, so nobody
worked on it.
On the business side, ATM and Ethernet have completely different
economic models. Ethernet is an open, competitive, multi-vendor
environment: you can plug things in; they're cheap; and people
understand how they work. ATM is basically a vertically integrated
proprietary network model: if you buy one ATM switch from one vendor,
it's not going to work with a switch from the next guy. And there is no
Bottomline, the fact is that ATM never worked as advertised. We have not
found one production installation that has deployed ATM in the way it
was intended. People have done all kinds of prototype experiments.
Everybody had to buy a switch, but in the end,the promise didn't get
filled. The technology did not deliver.
THE HERRING: What do you think was the impetus for ATM?
Bechtolscheim: It was the phone companies trying to redefine data
networking. The idea was to have separate protocols for data, voice, and
video. One key argument for the end-to-end connections was so they could
count the bits passing through and charge people for it. But the whole
development was completely decoupled from the Internet, and that was its
downfall. And calling IP the legacy protocol was the ultimate insult.
THE HERRING: Why didn't the creators of ATM anticipate this?
Bechtolscheim: Five years ago, there was no World Wide Web, so people
didn't foresee how dynamic the IP environment could be. The same is true
for IP Multicast. There are four different ways to run IP over ATM, none
of which works particularly well. It's completely hopeless.
Customers wanted high performance, scalability, and multimedia, but
there was no critical review of the underlying technologies. It's only
now that people are realizing that it doesn't matter because ATM is
THE HERRING: so you saw that a year ago?
Bechtolscheim: Yes. People in the industry were quite frustrated. Take
Cisco, which has made tremendous investments in ATM and hasn't made any
headway. And then there are Bay Networks and 3Com. They all promised
that ATM would unify the existing networks, Token Ring, FDDI, and
Ethernet in a big cloud. But it never was delivered. You can only keep
the hype alive for so long. At some point, you have to say, "It doesn't
work. Let's solve the real problem."
THE HERRING: But these companies have so much expertise in networking.
Bechtolscheim: The frame of reference was Cisco routers and hubs. This
whole switching thing is a very new phenomena. It only took off last
year. By the way, our tagline is "The Switch Is The Network." The
network is actually collapsing to a switch, meaning the new technology
is switching. You have to do switching to scale performance. And the
key challenge is combining switching with routing. That's the holy grail
THE HERRING: Is there a ceiling to Gigabit Ethernet?
Bechtolscheim: It could go 10,000 megabits. Right now, the optics beyond
1,000 megabits are very expensive. There is no volume market for high
speed optics technology. Plus, 1,000 megabits is roughly what you can
get into a host computer through a high performance PCI bus.
You can also be too early. Ultranet, had a gigabit network back in 1984,
but it was too fast for the machines back then. What's really happening
is that the network is just catching up with the speed increases of CPUs
over the last couple of years. It's very difficult for the network to be
a bottleneck to the CPU, but now, all these Pentium and Unix chips are
going faster than the network can perform.
With gigabit speed, the network is slightly ahead of the CPU, so the
network is no longer the bottleneck. And the funny thing is, it's not
that hard. You just make it go faster. People thought it was a hard
thing to do, but in reality, you just take some fiber channel optics
from the fiber channel effort and an Ethernet interface and re-implement
it in a higher speed logic. It's gigabit on a megabit. It doesn't even
cost that much more. There's no real cost associated with the speed
increase. It doesn't take any more gates.
THE HERRING: What are some of the challenges?
Bechtolscheim: Historically, these networks were built with hubs and
routers. Cisco makes the routers; Cabletron, Bay Networks, and 3Com make
the hubs. It was a simple picture. But both hubs and routers are
essentially bottleneck--or bandwidth--reduction devices. They're not a
suitable building block to build a scalable network. And at gigabit
speed, routers can't keep up. You basically have to go to a fully
switched network environment. You have to do the packet forwarding in
hardware to keep up with the millions and millions of packets that are
going to fly through there.
The industry hasn't invested the time to really pursue this. The whole
system employs a software approach that runs on some fancy CPU designed
to handle 250,000 packets a second. But a single gigabit wire contains
more than a million packets a second.
So it was really the challenge of scaling that attracted my interest.
And it comes down to how you switch Ethernet traffic at these high
speeds effectively at the network layer. You can't just get rid of the
routers; you have to do some of those functions. I guess I can't really
talk much more about what we exactly do there. But that's the
challenge--to do the right kind of switching in the network at these
high speeds. Because without this structure, you don't have a scalable
THE HERRING: Is your success linked to the development of gigabit
Bechtolscheim: We're doing that too, actually. We don't want to be in
this as an end-user business, but it's very important to us that there
is a large number of gigabit plugs out there, because we plug into them.
And the existing Ethernet NIC architectures are hopeless: they do not
scale up to gigabit performance; they weren't designed for high
performance. So we took some of our system designing expertise and
applied it to this problem.
THE HERRING: And those cards will end up being as cost-effective as ATM?
Bechtolscheim: They will be much cheaper than ATM cards.
THE HERRING: How are you planning on marketing your product?
Bechtolscheim: There are all kinds of companies that want to be in the
network space. There are the traditional networking companies, and then
there are the traditional system companies. They are both very capable
of supporting these kinds of customers.
Our product appears to be a particularly good match for an OEM business
model. We're planning to have a product out in the second half of
calendar '97. But that includes a lot of testing and development. You
have to be compatible with a lot of things out there, and you have to
get it certified and approved, and so on. We will have silicon on many
of the things we're doing this calendar year, but it will takesome time
until we actually ship.
Also, our product has a lot of software functionality. As most people
have learned, software takes even longer than hardware. Frequently, you
can only fully test it once you have the system all together. We are
allowing for that, as well.
THE HERRING: How much money have you invested?
Bechtolscheim: I initially committed $5 million to the company. We are
in the process of doing a corporate financing with our first OEM
partner. We are also working with our friends at Kleiner Perkins. But
at this stage, we don't need any more phone calls from people who want
to invest. You can print that in your article.
THE HERRING: What role does luck play?
Bechtolscheim: I don't think it's really luck so much as timing, meaning
doing the right product for the right market at the right time. That's
always the best. When Sun started, we had great timing. We could have
not started the company one year earlier or one year later. It just
wasn't there. I see the same thing here.
If we had started our company a year or two earlier, we might have ended
up in the ATM sandtrap; sometimes, people pick the technologies that
surround them. I don't think our timing could be any better from a
technology and market side. We still have to finish our product, but
that's under our own control.
In an article entitled "ATM Changes Everything" in THE HERRING's April
issue, Todd Dagres of Battery Ventures sang ATM's praises and claimed
that ATM "has the potential to be a revolutionary new networking
technology." Naturally, we were curious to find out what he thought
about some of Mr. Bechtolsheim's claims.
THE HERRING: What do you think about Mr. Bechtolsheim's statement that
there are four standards for ATM, none of which work?
Dagres: The argument of standards is a difficult one for him to make,
because right now, there are basically 16 versions of gigabit Ethernet.
ATM is much further along in terms of standards and the technology being
used. Granted, gigabit Ethernet is just starting, but his argument is,
"It's the Ethernet we know and love." If it's the Ethernet we know and
love, why are there are 16 different flavors of it?
THE HERRING: That brings us to another point. Mr. Bechtolsheim states
that he is building on an infrastructure that already exists. How would
Dagres: They're playing a bit of a marketing game by saying, "Hey, it's
Ethernet. All we are doing is turning up the clock speed." There's a new
physical layer called Fiber Channel, and then there are multiple flavors
for what you are going to place between the gigabit Ethernet MAC and the
Fiber Channel. And they haven't even agreed on what the gigabit Ethernet
MAC is going to look like.
THE HERRING: What do you think about the importance of TCP/IP?
Dagres: He focuses a lot on IP and says that we have to build networks
for IP. I disagree. I think we have to build networks to satisfy
application requirements not necessarily a given protocol. IP is
becoming dominant; I do agree with that. But I also think there are new
multimedia applications coming down over the Internet and corporate
networks that are not necessarily going to be IP-oriented.
THE HERRING: Isn't IP important?
Dagres: Ethernet is connectionless. It's this big shared media. For
certain applications, that's nice. If you want to build a high
performance multimedia backbone, I would argue strenuously that ATM is a
superior technology to gigabit Ethernet.
THE HERRING: Why?
Dagres: Ethernet is a very promiscuous technology; ATM is more secure.
Gigabit Ethernet is also going to have limitations in terms of distance
and how manageable it's going to be in a large backbone. If you're
going to build a gigabit Ethernet network, you're going to build a
bridge network. And we learned a long time ago that large, flat networks
don't work very well ATM looks like a flat network, but it actually can
act as a more hierarchical network based on the way that you build it.
THE HERRING: What do you think about his tagline that "The Switch Is The
Dagres: I think what he's saying is he doesn't want to build a gigabit
router, because that would be very expensive and add a lot of latency.
What he's also saying is that Ethernet in and of itself is not smart
enough and doesn't perform well enough to handle today's applications.
So what we have to do is build a switch with a lot of software and
intelligence built into it so that we can take the inherently dumb
Ethernet and make it act the way we want to. What it all comes down to
is we are gerry-rigging Ethernet to support very high speeds and
multimedia. That's why they're using Fiber Channel--to try and raise
Ethernet to ATM's level.
THE HERRING: So you think it will be difficult for gigabit Ethernet to
support multimedia applications?
Dagres: In the Ethernet market, you're sharing bandwidth. With full
duplex switched gigabit Ethernet, if you dedicate a switched link to an
application and you bury it with bandwidth, you can probably take care
of it. But what I'm saying is that ATM is an inherently more efficient
multimedia supporting technology. With gigabit Ethernet, the aim is to
try to beat the problem of differing types of service quality just by
adding bandwidth and by switching it, whereas ATM is inherently a
switched technology and, thus, inherently superior for multimedia
THE HERRING: What is gigabit Ethernet good for?
Dagres: I think gigabit Ethernet will be successful as a connection
medium for connecting high speed servers to a backbone and for
connecting 100 megabit workgroups to a backbone. As a backbone
technology in and of itself, it could work well in a collapsed switch or
collapsed backbone environment.
THE HERRING: But you do not think that it will supplant ATM?
Dagres: I am not a total ATM bigot, but to say that ATM is dead is
invalid, because customers are deploying it in their backbones. I agree
that ATM to the desktop has been disappointing, because people don't
need 155 mbps to the desktop, and the advantage of 25 mbps to the
desktop is not a quantum leap in capability from 10 mbps switched. But I
do think that ATM has proven to be a very valid technology for the
Real human wealth, in the form of security, freedom, productivity, and
knowledge, is scarcely captured by unexercised stock options. There are
no line items that gauge the real engines of prosperity: vision,
passion, and commitment. The plain truth is you cannot suck reality
from the hypnotic glow of a vacuum tube.
-- Jerry Kaplan