TECHNOLOGY REPORT

Published jointly by Gilder Publishing, LLC and Forbes Magazine www.gildertech.com February 2001/Vol. VI No. 2

The Storewidth Warp

Exhausted by the travails of time and space, Charlie Burger’s jet-lagged body
craved the San Jose Hyatt. Burger as you know is our main man on optics, and a key
fact of optics is that photons cannot be stored. They are pure moving energy and by
BlueArc nature cannot be settled in a capacitor or trapped in an insulated gate like an elec-
imperils all tronic charge in a silicon memory, or durably defined like a magnetized domain on
a disk, or deposited comfortably at a Hyatt or a Motel 6, even if management leaves
the software a light on for them. Although Science magazine recently recounted experiments in
based net- which photons were slowed down and even stopped for a millisecond or so in a lab-
work storage oratory, a close reading revealed that in the process the photons died.
Charlie is a lot like a photon. But he also has some 155 pounds of mass,
appliances, honed to a hard heft by a regimen of slow power lifting. After 18 hours of air-
whether it be ports and rental cars and interviews on optics he needed to stop for somewhat
more than a millisecond and consign himself to a good night’s storage, even if
from NetApp, he had to undergo a discrete cosine transform and curl up compressed on a 19
EMC, HP, inch rack in an Exodus (EXDS) data center.
But it was not to happen. Before Charlie’s head could hit a pillow, he would
Procom or be abducted by aliens and taken on a trip into the storewidth warp, where he
Compaq would discover what may prove to be the most disruptive company in the histo-
ry of this letter. Within a week, as a freak effect of global warming, snow would
fall in Silicon Valley and the entire storewidth landscape would be frozen in
anticipation of a BlueArc from outer space.
Living the time-space dilemmas of storewidth, Charlie was torn between optics and
storage. A time-bound activity, storage enables movement of information from one
moment for use in another. The bits go somewhere and stay put until they are con-
sumed serially by time-bound humans. Communication is space travel. It moves infor-
mation through space, from one place to another.
In This Issue: With many transmissions possible between the same two
points at once, communication ideally is parallel and real time.
Cover Story: Tredennick’s paradigm; The ultimate parallelism is wavelength division multiplexing
Avanex accessory to kidnapping; Silicon (WDM)—putting data beams of many colors at once on a single
storewidth; The BlueArc revolution; fiber thread and moving them across the land at lightspeed.
Altera’s storewidth link; Mango magic Between time and space is the storewidth warp.
In the scheme of abundances and scarcities that shapes the
Disk Performance Degradation Page 4
economics of technology, all computing can be mapped to this
Telecosm Table Page 8 matrix of time and space. The usual computer architecture
 embodied in the typical Pentium is a time machine. With Crucial to overcoming the bottleneck between storage
over 90 percent of the chip devoted to various forms of and fibersphere–between time-bound electronics and
memory, the actual computation uses a tiny portion of spacious optics, between you and the Web–is the devel-
the chip. But it uses this tiny space repeatedly at a fren- opment of space machines on the edge. Needed are
zied pace to execute a series of software instructions on storewidth servers that can match the parallel flood of
a flow of data words fetched from memory. It cuts time photons with parallel paths of electronic processing.
into nanosecond slices—billions of cycles every second— Relinquished must be some of the general purpose flexi-
and if Intel’s (INTC) heroic chip designers have their bility of the microprocessor. But today the microproces-
way, two gigahertz Pentiums will soon slice time into a sor era is ending. All technologies must adapt to the awe-
few hundred picoseconds (trillionths of seconds). some power of the fibersphere.
Emblazoned in this architecture is the idea that space
is scarce. By spreading out computations over time, seri- Avanex accessory to kidnapping
al processing economizes on space. All calculations are Before Charlie encountered his storewidth epiphany,
done in one place, the central processing unit (CPU), he gained new appreciation for the power of this optical
one after another in time. Since only one calculation can technology that was rapidly transforming the Peninsula
occur at any moment in time, speed is gained by short- into a Silica Valley of glass. Capping Charlie’s day,
ening the moments. Simon Cao of Avanex (AVNX) revealed his plans for a
spectacular presentation of the Metro PowerExpress
Beset by ever lengthening TDM user and PowerExchanger at the March 2001 Optical Fiber
Conference (OFC) in Anaheim. The seamless metro-
queues on the network side, web politan area network would cut the fiber ribbon, so
servers face the demands of hetero- Simon implied, on an era of networks with thousands of
parallel lambdas spread across the Web without any
geneous data on the storage side switches in sight.
How could networks do without switches? Charlie
Parallel processing, by contrast, economizes on time by envisaged long lines of Cisco (CSCO) and Nortel (NT)
spreading computations out over space. Rather than con- engineers jumping off the San Mateo Bridge. Networks
ducting the operations in the CPU, the CPU stands to the almost by definition had to be switched. But switches
side and controls the pattern of links through which the data imply buffering and storage along the path and thus
flows. A data flow machine resembles a piano, with the intolerable delays on the network itself. Simon’s wave-
piano wires representing the logic gates and the keys repre- length division multiplexed networks would ride lambdas
senting the control inputs. As Carver Mead and Lynn (wavelengths) that could be added or dropped optically,
Conway put it in Introduction to VLSI, “A complex function muxed and demuxed, without an actual cross connec-
may thus be performed...by the data path just as the static- tion. At the beginning of the trip you choose your lamb-
appearing array of piano wires may produce a complex and da and the light does the rest. If you had fiber every-
abstract piece of music when a series of notes and chords where, tunable lasers, filters, couplers, and splitters, and
are struck in a particular order.” a lambda for every route, it could work. If I had any
bread, I could make a ham sandwich, if I had any ham.
Tredennick’s paradigm Undaunted by the admittedly long run challenge of
The fibersphere is such a space machine. Wavelength removing switches altogether, Cao is launching his con-
division multiplexing is inherently parallel and pianistic. You cept in the metropolitan area network where it is already
can’t store photons. As wavelengths multiply, they fly in par- taking hold. The campaign could take a decade. Perhaps
allel through the fibersphere. By proliferating lambdas–opti- a switch or two could be sold in the interim. (The optical
cal pathways or circuits–photonic networks enable many switch guys could come in off the bridge).
users to share the same fiber pathways yet still have their In any case, absorbing the implications had over-
own exclusive circuits. In the fibersphere there is no time loaded Charlie’s processor for the day, and he had begun
sharing or time division, no TDM process of distributing the his descent somewhere over the WDM rainbow. In the
data packets of many users in time slots on the circuit. parking lot his Dollar Rent-a-Car reflected the long seven
As the fibersphere explodes into ever more parallel paths, hundred nanometer rays of the California dusk. The
giving more and more users access to circuits, computers at day’s final hard drive beckoned.
the edge can no longer keep up by slicing time into smaller Easy to please, Charlie would have settled for simple
moments. Even as massively parallel optical bandwidth and stupid highway attached storage, without queues and
overwhelms the TDM and SONET electronic switches in buffers in the lobby or tricky transactional protocols at
the core of the network, massively parallel throngs of users the front desk. He was ready to go without guaranteed
jam the serial electronic servers on the edge. Already beset quality of service or redundancy (one bed would do) and
by ever lengthening TDM user queues on the network side, without chocolate wafers on his pillow, data masseuses
web servers face the increasingly complex demands of het- on his disks, or 50 millisecond restoration if the
erogenous data on the storage side. California lights browned out.

2 GILDER TECHNOLOGY REPORT

 But what’s this? As he walked toward his car, he file servers that deliver web pages and other program-
became aware of a distraction. “No,” demurred Avanex’s ming to the Net. A web server is a robust general purpose
VP of investor relations, Tony Florence, “over here.” He computer that can operate 24 hours a day on any of thou-
pointed to a limo with a chauffeur in a black suit and sands of different software programs devoted to commu-
dark sunglasses. Gently ushered into the black Lincoln nications protocols stacks, authentication and security
Towncar, Charlie heard Tony’s last words as the door rules, file transfer protocols, messaging and transactions
shut: “Ask for Pesatori, code word Alessandrini.” schemes, all retrieved from locally attached disks.
Straining to see the dim lights through the tinted win- Until recently servers were actually faster than the
dows, Charlie asked his driver where they were headed. network, so you didn’t have to worry about them. But
“North, I believe.” with the arrival of WDM in the backbone and gigabit and
At last Charlie caught a glimpse of a sign. “South” was multi-gigabit Ethernet optics in the metro, campus, and
all he could make out. Forty minutes later he found him- LAN, storage file server speeds have become the Web’s
self in the parking lot of what appeared to be a deserted major sticking point.
office building. Definitely not the Hyatt. He knocked at
the entrance. The faster Ethernet became, the
From behind him came a sharp, “I’ll wait here. You
may need me.” Generous, he thought, of his dour driver. more of a bottleneck the NAS
In the darkened foyer, one of what seemed to Charlie to device in its path turned out to be
be Earth’s last two employees opened the locked door.
Charlie tried to summon the code. “Pesadrini, The bottleneck comes less from the processor itself,
Allesantori…. Synaxia… no BlueArc… but don’t tell any- which has moved over the last decade from scores of
one…. Whatever.” His codewords had no effect. Charlie megahertz to the gigahertz, than from the internal buses,
turned on his heels to leave. which move data bits around the computer backplane for
“Stop!” processing. Since 1993, Pentiums have connected to
Charlie froze mid-step. memory through the PCI (Peripheral Component
“Down the hall. They’ve been expecting you.” He fol- Interconnect) bus, which has increased its speed merely
lowed his guide through the empty corridor into a dimly- from 100 megahertz to 133 megahertz during that period.
lit office. On the right, three men encircled an imposing Bandwidth inside the computer has fallen behind
table. “Enrico Pesatori—friend of Walter Alessandrini,” bandwidth outside the computer; outside optical band-
one announced himself in a deep, old country voice. width is growing some ten times faster. It was superior
“We’re in stealth mode,” he said, superfluously. Charlie input-output (I/O) speeds that put the computer in a box
glanced back for his chauffeur. The man was gone. The with its storage in the first place. As optical network
door clicked, and a revolution in storewidth unfolded. speeds destroy the same box advantage, the computer is
disaggregating across the Net. Since a computer is
Silicon storewidth essentially an I/O device for accessing and processing
Whether under its original name Synaxia, or the con- storage, the movement of storage to the Net hollows out
sultant’s new moniker BlueArc, the storage devices on the computer and makes way for BlueArc.
the edge of the network would at last conform to the nat-
urally parallel optics in the center. Matching the paral- The BlueArc revolution
lelism of the fibersphere would be a parallel server- It turned out that BlueArc had nabbed the wrong GTR
sphere. In the storewidth warp, time and space would analyst. Charlie arrived back in Housatonic, eyes glazed
converge. The Internet would never be the same again. over, mumbling something about tinted windows and
Hey, you’ve heard that before, when Kim Polese first sequential Iometers, insisting we call a gentleman named
shook her Castanet at Telecosm, when Amazon began Enrico. Determined to get to the bottom of things, we put
giving away “Champagne smoochies soap and aromatic in a call to this Enrico, and set up a meeting.
votives” on its website, when Pointcast launched the One week later, in New York at the Gerson Lehrman
vaunted age of “Push” in Wired, and when Matt Drudge Group offices in Manhattan, which provide crucial research
levered the political landscape from a trashcan on the to the GTR through their huge database of technical
Web. You heard it too when digital subscriber line (DSL) experts, we finally met Charlie’s storewidth angels. Present
and cable modem technologies were unleashed to bring was not only CEO Enrico Pesatori, but also Dr. Geoff
broadband access to the Web for all, guaranteed by prom- Barrall, CTO, of the BlueArc Corporation. Neither Charlie
ises from politicians and commissioners in Washington nor I had any clear idea what a storewidth bonanza he had
and the fifty states. You heard it before. Now it is truly uncovered during his magic carpet ride in Silicon Valley.
happening–and none too soon… Two years ago, working to improve Ethernet switches,
Even with a broadband local loop in place, the Net Dr. Barrall, a Ph.D. in Cybernetics, and his colleagues in
will encounter a bottleneck in the persistence of old, a U.K.-based consultancy quickly discovered that success
time-bound computer architectures in the face of the was fruitless. The faster Ethernet became, the more of a
space travel of optics. The result is a jam at the storage bottleneck a network attached storage (NAS) device

FEBRUARY 2001, VOLUME VI NUMBER 1 3

 DISK PERFORMANCE DEGRADATION
Exponentially mounting data supplies have driven the demand Disk storage prices have declined by a
for storage capacity through the roof. Historically, disk manufac- factor of 30,000 in the past 40 years.
turers have heroically met the challenge of providing consumers Chart 1
with smaller, faster, and cheaper forms of disk storage. $350
In the past five years, magnetic disk storage has experienced its
steepest price decline in forty-three years (Chart 1), making mag- $300

Cost per Gigabyte of Magnetic

Disk Storage (U.S. Dollars)
netic disk storage the data storage method of choice (Chart 2). $250
In the early 1990’s disk read/write heads soared more than 90
nanometers (billionths of a meter) above the disk’s surface (Chart $200
3). Head technology improvements over the past decade have
$150
brought the physical distance between the head and the disk sur-
face down below 10 nanometers. Advanced head technologies $100
have led to bit cell size and disk trackwidth reductions (Chart 4).
Smaller, faster, cheaper inescapably leads to denser, smaller $50
bits on narrower tracks. Decreasing the size of a bit cell by a 0
factor of two quadruples disk density. Areal density, the prod- 1995 1996 1997 1998 1999 2000
uct of track density and the number of bits per track, has
increased in by a factor of three million in just over four
decades (Chart 5).
Disk performance, however, has not maintained the
Data supplies have driven storage
Moore’s Law-pace of disk densities. Densely packed bits mean sales through the roof.
3000 Chart 2
fewer disks for a given amount of storage and thus fewer drive
heads per gigabyte. Time to access a gigabyte, therefore, is
Total Disk Storage Capacity Sold
2500
actually deteriorating as seek times improve at a relatively slow
linear rate while densities surge exponentially (Chart 6).
2000
Disks have become a performance bottleneck. High per-
(petabytes)

formance alternatives to magnetic disk storage have a strong 1500

growth potential. Solid-state disk storage from companies like
Solid Data and Quantum (HDD) addresses the disk drive I/O 1000
bottleneck head-on by caching small highly transactional data
in a separate cache to improve overall disk drive performance. 500
InPhase Technologies, launched last week out of Lucent’s Bell
Labs, will address the magnetic disk bottleneck by focusing on 0
holographic storage. 1995 1996 1997 1998 1999 2000
-Mary Collins
Data Sources: IBM, International Data Corporation

placed in its path turned out to be. Network attached file per second PCI bus (I/O interface), for example, can slow to
servers had not yet entered the optical age. a standstill while main memory and CPU handle one cus-
High-speed Ethernet demanded high-speed storage tomer at a time. They step and fetch software instructions for
servers. Yet all existing NAS appliance designs were based on the Internet’s TCP/IP (Transport Control Protocol/Internet
traditional server architectures. Designed for computation, Protocol) stack, then shuttle back and forth from memory to
they become I/O bound in today’s high throughput networks. carry out Windows and Unix file transfer protocols.
Two gigabit Ethernet (2,000 Mbps) coming in off the LAN Upgrading your NAS from a one thousand dollar system to a
and 1.7 gigabit per second fibre channel at the storage end one hundred thousand dollar system buys a lot more storage
swamp even the fastest storage servers on the market, such as capacity but affords essentially the same throughput.
the NetApp 840 and the EMC IP-4700. These software Barrall got together with Pesatori, a former Digital
intensive devices traditionally run at 200 Mbps to 300 Mbps, Equipment and Compaq (CPQ) VP and president of Tandem
400 Mbps when streaming audio or video. Computers (now part of Compaq). Together they assembled
Under Amdahl’s law, high speed server components can a team of sixty engineers to build a multi-gigabit throughput
operate no faster than the slowest link on the data path. The server from the ground up.
whole is not better than the sum of the parts. A one-gigabit A server capable of reading and writing simultaneously at

4 GILDER TECHNOLOGY REPORT

 Disk drive head technologies Disk densities have increased at a
continue to improve... Moore’s Law-pace, contibuting to...
100 100,000
Chart 3 Chart 5
90
Spacing Between Reader Head and

10,000
80
Disk Surface (nanometers)

1,000

Areal Density (Mb/inch2)

70
60 100
50
10
40
30 1
20 0.1
10 0.01
0
1993 1994 1995 1996 1997 1998 1999 2000 0.001
‘57 ‘61 ‘65 ‘69 ‘73 ‘77 ‘81 ‘85 ‘89 ‘93 ‘97 ‘01

... driving bit cell and trackwidth

size down exponentially. ...disk performance degradation.
18 Chart 6 20,000
12 Chart 4 0.6

10 16 16,000
0.5
Bit Length
Seek Time (milliseconds)

Areal Density (Mb/inch2)

Trackwidth (micrometers)

Bit Length (Micrometers)

Trackwidth 14
8 0.4 12,000

6 12
0.3
8,000
4 10
0.2
4,000
2 8
0.1

0 6 0
0
1990 1992 1994 1996 1998 2000 1990 1992 1994 1996 1998 2000

multi-gigabit rates could not be built with off the shelf com- age protocols, with all their hundreds of lookup tables, direc-
ponents. Dr. Barrall and his team would create all server tories, and connection lists, in hardware. One of the parallel
components from scratch. Born was the first Silicon Server modules contains the world’s fastest all-hardware TCP/IP
architecture, operating at wire speed between gigabit networking subsystem. Another holds the world’s fastest
Ethernet at the network side and gigabit fibre channel at the fibre channel controller. Sandwiched in the middle are
storage end, with drastically fewer components and vastly diverse protocol translation units and file subsystems that
higher performance. perform all the key storewidth functions–SAN, NAS, you
Suggestive of EZ Chip’s (LNOP) pipelined chip architec- name it–of delivering data from storage to the Web.
ture, with TOPs (task optimized processors) each designed to BlueArc initially offers a single box with throughput of 2.5
perform one specialized function, BlueArc’s Silicon Server gigabits per second between gigabit Ethernet at one end and
architecture separates and pipelines the four key server func- gigabit fibre channel at the other. Scaling up to 200 terabytes
tions into four parallel processing modules. Designed more of storage through a single Ethernet connection, BlueArc
like a switch or a router than like an ordinary server, the blows away the sixteen terabyte limit of current NAS and
BlueArc device blasts open the bottleneck between users and SAN devices. Consuming nearly half the power and scaling
storage by putting the TCP/IP stack, the file servers, and stor- to tens of thousands of simultaneous users, BlueArc pulver-

FEBRUARY 2001, VOLUME VI NUMBER 1 5

 izes the current server limits of less than two hundred. Between 1985 and 2000 circuit density rose one thou-
For 20 percent of the capital costs, this architecture sand fold while processor speed increased by a factor of
offers five times the bandwidth for streaming applica- twenty. Field programmable gate arrays exploit this new
tions, 10 times the bandwidth for traditional file trans- abundance by distributing processing functions across mil-
fers, 30 times the amount of accessible storage, and a lions of gates operating in parallel across a chip. Such par-
hundred times the number of simultaneous user connec- allelism is usually feasible only when the processing func-
tivity, all with five-9s solid state reliability. It is a “Silicon tions are hard wired, limiting flexibility, but FPGAs can be
Server for the optical age.” “rewired” in milliseconds. With Xilinx and Altera introduc-
Think of the Silicon Server as a data flow processor ing arrays with as many as 10 million gates or logic ele-
between the time-based world of storage and the space- ments (comprising some 100 million transistors) this year,
based world of the fibersphere. To the network, it looks like increased chip density makes the FPGA an effective rival to
any other Network Attached Storage device. With the tra- the microprocessor for many applications that do not entail
ditional server bottlenecks removed it achieves multi-giga- a strict temporal order of execution.
bit throughput. It supports all relevant storewidth proto- Since a storage server primarily directs the data flow,
cols and can be dropped in to replace any other device that a spatial function, rather than processes it, a temporal
was previously occupying that space. Ultimately gone are function, FPGAs are the natural solution. Just as optical
all the software based Von Neumann network appliances, switches outperform electronic switches by merely steer-
whether from NetApp (NTAP), EMC (EMC), Procom ing data paths through space rather than processing
(PRCM), Hewlett Packard them in time, parallel FPGAs use
(HWP), or Compaq. With unified millions of gates to steer data
FPGA Costs
loads of 200 terabytes and tens of across the chip rather than forcing
thousands of clients, gone too are $350 it into time queues to be processed
all the load-balancing gear from $300 in the CPU.
Cisco, Alteon, Extreme Networks In the BlueArc box, which uses
Cost per 1 Million Gates

(EXTR) and other stalwarts of the $250 Altera devices, FPGAs perform
network storage center. functions in the spatial domain,
BlueArc’s secret is the para- $200 such as spatial data flow, that were
digm. In the past, with protocols, $150 previously forced to conform to a
file stystems, backup schemes, and temporal processor. Altera Apex
I/O technologies in constant $100 FPGAs perform all hardware data
change and turmoil, all such func- transport with mere hundreds of
$50
tions had to be performed in soft- CPU MIPS (millions of instructions
ware. BlueArc recognized that the 0 per second) rather than the thou-
latest generation of configurable 1998 1999 2000 2001 sands consumed by software inside
field programmable gate arrays of traditional servers. FPGAs give
(FPGA) could execute these same tasks in hardware with the Silicon Server a true performance advantage, one that
orders of magnitude improvements in speed, and still be would not have been possible a year ago and will continue
reconfigured in less than one minute, even milliseconds, to grow over time, as costs continue to decline.
when new technologies emerged. FPGAs have been a FPGAs excel in functions–encryption/decryption, pat-
favored GTR technology from the beginning and are now tern recognition, and real time signal processing–that are
being celebrated and explained in depth in our new critical on the Net. In order to move storage to the net-
Dynamic Silicon letter written by Nick Tredennick. work, it must be encrypted. In order to communicate
Represented on our Telecosm list by Xilinx (XLNX), across the network, signals must be constantly processed
Atmel (ATML) and now by Altera (ALTR) as well, these in real time. Speech recognition, optical character recog-
configurable devices feed on another of the time-space nition, error correction, code division communications,
differentials that will shape the future of the industry. and image capture all entail pattern matching.
Transforming the file server into a bi-directional data
Altera’s storewidth link flow engine, BlueArc is the first company fully to exploit
Microprocessors tend to economize on space, and the spatial resources of field programmable gate arrays
waste time. The actual processor occupies less than one and channel all data flow inside the hardware.
tenth of the device; yet the standard von Neumann archi- In the presence of Dynamic Silicon author
tecture dictates that the chief way to augment perform- Tredennick, BlueArc ran its first generation Silicon
ance is to use this same small silicon area more and more Server through such diverse performance metrics as
millions of times a second to execute a series of instruc- Iometer (Intel storage benchmark), Spec SFS (UNIX),
tions on data fetched for it from off-chip memories. Yet and Netbench (Windows). “They let me look into the
even as this processing time remains precious and scarce, box. They let me change the parameters any way I want-
processing “space”–the density of circuits on a chip–has ed. Nothing stopped the machine from working at its
become relatively abundant. peak rates.” Not a general purpose server, it does not exe-

6 GILDER TECHNOLOGY REPORT

cute the thousands of server functions, from searches to 2001. Under the guidance of CTO Scott Davis, the com-
messaging to graphics rendering that any general purpose pany has launched two products–Cachelink, which uses
Unix or Solaris box from Sun (SUNW) or NT machine the browser cache in your computer to accelerate access
from Compaq can perform. But BlueArc is the first net- to favored web pages in a company, and now
worked storage appliance that addresses the show-stop- Mangomind, which provides storewidth file services
ping throughput bottleneck imposed by traditional server across the Internet.
architectures. It breaks through the storage bottleneck With the ultimate edge advantage, close proximity to
on the Web and reconciles the time domain of the Web the end user, Mango has accomplished what highly dis-
with the outer space of the fibersphere. It is a machine tributed content delivery networks have attempted for
for the storewidth warp. We put them on the list both as years. Mango is as close to the edge of the network that
a potential private investment and a likely public issue. you can get–on the users’ desktops. Mango has effec-
tively transformed the desktop into the smartest, most
Mango magic efficient cache on the network, a cache that knows exact-
Virtual private networks (VPNs) are dedicated and ly what to cache, when, and for how long. Cache
encrypted paths that enable security conscious corpora- coherency becomes a non-issue. Files are cached and
tions to send mission critical files over shared data net- content verified using Mango’s version check protocol
works such as the Internet. Mangosoft (MNGX.OB) each time an end user opens or saves a file. Files that
does not belong to the VPN Consortium, nor does have been altered by users working offline are transpar-
Mango manufacture VPN hardware. Mango does, how- ently synchronized when the user reconnects to the Net.
ever, have Mangomind, a LAN-style file service that No extra time, IT power, or CPU power is wasted main-
enables VPN-like applications across the Internet. taining cache content freshness.
Mangomind extends the client’s existing network servic-
es by allowing secure storage and sharing of data Hollowing out the computer,
between members of extended business workgroups.
While traditional VPNs require dedicated phone lines Mangomind uses a distributed shared
and dedicated IT staff to collect user profiles of remote
peering partners and to install, configure, and manage
memory architecture originated for
VPN hardware, Mango was able to set Gilder massively parallel supercomputers
Publishing up with our very own VPN and remote
Internet storage cache with a single e-mail. Most storewidth companies pay top dollar for storage
The first member of a Mangomind drive can invite IT professionals to maintain network efficiency and
others to join by e-mailing client software for download- cache coherency on their content delivery and storage
ing on their PC. Each member contributing files and networks. Mango, in an ironic twist, doesn’t pay, but is
folders to the drive can regulate access of other drive paid by its systems administrators. Given that close to 90
members to them. Multiple users can simultaneously and percent of the business world is proficient in Windows
securely access and edit Mind drive files. Use requires no and that Mangomind was designed with bug-for-bug
training, no additional hardware, and no IT intervention. Windows compatibility, the end user and the cache
The Mind drive does all of the thinking, arbitrating file administrator are one in the same. Although Mango is a
reads and writes. All applications operate directly on files very thinly traded stock, we put it on our list with the
without creating multiple incoherent copies. usual cautions. Among the board members is Nick
Files saved to the Mango drive are encrypted at the Tredennick of Dynamic Silicon.
end user prior to transmission over the Internet to
Mango’s centralized network cache. Managed by Storage – George Gilder with Mary Collins
Networks (STOR) and located at the Exodus data center February 13, 2001
in Boston, the Mango cache stores data in encrypted
form. By contrast, a traditional VPN scheme decrypts
files at the destination storage server. In the Mango sys- Inserted with this issue is an illustration of The
tem, all decryption takes place at the client–on the edge. Storewidth Landscape. On the back of the insert
Hollowing out the computer by combining the stor- you will find a list of players in the storewidth
age and memory of multiple users, Mangomind uses a space. Companies highlighted are participating in
distributed shared memory architecture originated for Gilder Publishing’s Storewidth 2001 conference.
massively parallel supercomputers at the late lamented
Kendall Square Research. Under the guidance of KSR This month the Gilder Technology Report’s
co-founder Steve Frank, that company developed sys- Telecosm Table has made the change from frac-
tems to share memory among many microprocessors tions to decimals. Also note that technology cat-
within its proprietary supercomputer. Frank then egory titles have changed, but companies within
brought the basic concepts to Mango, where they could those categories have not changed.
be executed in software on the accelerated hardware of

FEBRUARY 2001, VOLUME VI NUMBER 1 7

TELECOSM TECHNOLOGIES
ASCENDANT TECHNOLOGY COMPANY (SYMBOL) REFERENCE JAN ‘00: 52 WEEK MARKET
DATE / PRICE MONTH END RANGE CAP
FIBER OPTICS
Wireless, Fiber Optic Telecom Chips, Equipment, Systems Lucent (LU) 11/7/96 11.78 18.60 12.19 - 70.66 63.0B
Wireless, Fiber Optic, Cable Equipment, Systems Nortel (NT) 11/3/97 11.50 38.23 29.00 - 89.00 116.8B
Optical Fiber, Photonic Components Corning (GLW) 5/1/98 13.64 56.71 42.83 - 113.33 51.8B
Wave Division Multiplexing (WDM) Components JDS Uniphase (JDSU) 6/27/97 3.63 54.81 37 - 153.42 52.7B
Adaptive Photonic Processors Avanex (AVNX) 3/31/00 151.75 64.63 39 - 273.50 4.2B
All-Optical Cross-Connects, Test Equipment Agilent (A) 4/28/00 88.63 54.55 38.06 - 162.00 24.9B
Tunable Sources and WDM Components New Focus (NUFO) 11/30/00 20.31 60.19 16 - 165.13 3.8B
Crystal-Based WDM and Optical Switching Chorum (private) n/a n/a n/a n/a n/a
WDM Metro Systems ONI (ONIS) 12/29/00 39.56 55.50 22.25 - 142 7.3B

LAST MILE
Cable Modem Chipsets, Broadband ICs Broadcom (BRCM) 4/17/98 6* 109.94 72.38 - 274.75 25.9B
S-CDMA Cable Modems Terayon (TERN) 12/3/98 15.81 6.50 3.50 - 142.63 428.3M
Linear Power Amplifiers, Broadband Modems Conexant (CNXT) 3/31/99 13.84 18.06 12.63 - 132.50 4.4B

WIRELESS
Satellite Technology Loral (LOR) 7/30/99 18.88 5.90 2.69 - 21.00 1.8B
Low Earth Orbit Satellite (LEOS) Wireless Transmission Globalstar (GSTRF) 8/29/96 11.88 0.81 0.72 - 36.44 86.1M
Code Division Multiple Access (CDMA) Chips, Phones Qualcomm (QCOM) 7/19/96 4.75 84.06 51.50 - 162.56 63.3B
Nationwide CDMA Wireless Network Sprint (PCS) 12/3/98 7.19 * 30.50 17.63 - 66.94 28.4B
CDMA Handsets and Broadband Innovation Motorola (MOT) 2/29/00 56.83 22.81 15.81 - 61.54 49.8B
Wireless System Construction and Management Wireless Facilities (WFII) 7/31/00 63.63 40.88 27.13 - 163.50 1.8B

GLOBAL NETWORK
Metropolitan Fiber Optic Networks Metromedia (MFNX) 9/30/99 12.25 15.94 9.13 - 51.88 8.8B
Global Submarine Fiber Optic Network Global Crossing (GX) 10/30/98 14.81 22.02 11.25 - 61.81 19.5B
Regional Broadband Fiber Optic Network NEON (NOPT) 6/30/99 15.06 16.31 3.50 - 159.0 305.9M
Telecommunications Networks, Internet Backbone WorldCom (WCOM) 8/29/97 19.95 21.56 13.50 - 52.50 62.2B
Global Submarine Fiber Optic Network 360networks (TSIX) 10/31/00 18.13 14.50 10.0 - 24.19 11.8B

STOREWIDTH
Directory, Network Storage Novell (NOVL) 11/30/99 19.50 8.63 4.78 - 44.56 2.7B
Java Programming Language, Internet Servers Sun Microsystems(SUNW) 8/13/96 6.88 30.56 25 - 64.66 98.4B
Network Storage and Caching Solutions Mirror Image (XLA) 1/31/00 29 9.03 2.81 - 112.5 957.9M
Disruptive Storewidth Appliances Procom (PRCM) 5/31/00 25 21.31 10.25 - 89.75 247.6M
Remote Storewidth Services Storage Networks (STOR) 5/31/00 27* 27.94 16.50 - 154.25 2.6B
Complex Hosting and Storewidth Solutions Exodus (EXDS) 9/29/00 49.38 26.63 14.88 - 89.81 14.7B
Hardware-centric Networked Storage BlueArc (private) 1/31/01 n/a n/a n/a n/a
Virtual Private Networks, Encrypted Internet File Sharing Mangosoft (MNGX.OB) 1/31/01 1.00 1.00 0.75 - 28.0 26.9M

MICROCOSM
Analog, Digital, and Mixed Signal Processors Analog Devices (ADI) 7/31/97 11.19 62.60 42.63 - 103.00 22.4B
Silicon Germanium (SiGe) Based Photonic Devices Applied Micro Circuits (AMCC) 7/31/98 5.67 73.59 32.74 - 109.75 21.8B
Programming Logic, SiGe, Single-Chip Systems Atmel (ATML) 4/3/98 4.42 17.00 9.38 - 30.69 7.9B
Single-Chip ASIC Systems, CDMA Chip Sets LSI Logic (LSI) 7/31/97 15.75 24.75 16.30 - 90.39 7.9B
Single-Chip Systems, Silicon Germanium (SiGe) Chips National Semiconductor (NSM) 7/31/97 31.50 28.70 17.13 - 85.94 5.0B
Analog, Digital, and Mixed Signal Processors, Micromirrors Texas Instruments (TXN) 11/7/96 5.94 43.80 35.00 - 99.78 75.8B
Field Programmable Gate Arrays (FPGAs) Xilinx (XLNX) 10/25/96 8.22 54 35.25 - 98.31 17.8B
Seven Layer Network Processors EZchip (LNOP) 8/31/00 16.75 16.44 5.63 - 43.75 106.1M
Network Chips and Lightwave MEMS Cypress Semiconductor (CY) 9/29/00 41.56 27.33 18.25 - 58.00 3.6B
Field Programmable Gate Arrays (FPGAs) Altera (ALTR) 1/31/01 30.25 30.25 19.63 - 67.13 12.0B
ADDED TO THE LIST: BLUEARC, MANGOSOFT, AND ALTERA DELETED FROM THE LIST: C-CUBE * INITIAL PUBLIC OFFERING

NOTE: The Telecosm Table is not a model Gilder Technology Report Published by Gilder Publishing, LLC and Forbes Inc.
portfolio. It is a list of technologies in the
Monument Mills • PO Box 660 • Housatonic, MA 01236 FOR SUBSCRIPTION
Gilder Paradigm and of companies that lead
in their application. Companies appear on Tel: (888)484-2727 • Fax: (413)274-3030 • Email: [email protected] INFORMATION
this list only for their technology leadership, TELEPHONE
without consideration of their current share EDITOR: George Gilder
TOLL FREE:
price or the appropriate timing of an invest- PUBLISHER: Richard Vigilante
ment decision. The presence of a company
on the list is not a recommendation to buy
ANALYSTS: Charles Burger, Mary Collins, Bret Swanson, Jeff Stambovsky (888) 484-2727
shares at the current price. Reference Price MANAGING EDITOR: Debi Kennedy
is the company’s closing share price on the DESIGNER: Julie Ward Website:
Reference Date, the day the company was
www.gildertech.com
added to the table, typically the last trading SUBSCRIPTION DIRECTOR: Rosaline Fernandes
day of the month prior to publication. Mr. Copyright © 2000, by
Gilder and other GTR staff may hold posi- GENERAL MANAGER, GILDER TECH.COM: David S. Dortman
Gilder Publishing, LLC
tions in some or all of the stocks listed. PRESIDENT: Mark T. Ziebarth