杏吧原创

IT鈥橲 NOT the first time David Patterson has argued that computers are
carrying too much fat. In the early 1980s, when designers were cramming more and
more instructions into the vocabularies of their microprocessors, Patterson was
among the first to recommend reining things in. A computer should not be
verbose, he argued, but terse and fast. This bare-bones approach, called Reduced
Instruction Set Computing (RISC), revolutionised the field. Now, inside the
world鈥檚 fastest microprocessor, Digital Equipment鈥檚 Alpha chip, beats a RISC
heart.

Computers today have again become overblown, says Patterson, who is professor
of computer science at the University of California, Berkeley. If the
microprocessor鈥檚 future lies in battery-run, hand-held computers and memory-hungry
multimedia, then power-hungry chips like the Alpha and Pentium may become
endangered species. There would then be a niche for a new breed of stripped-down
microprocessor, a device Patterson calls 鈥渋ntelligent memory鈥� or IRAM. This
would combine a bit of logic and a huge memory on the same chip鈥攁 silicon
creature with the mind of a gnat and the memory of an elephant.

Patterson is not the only one who wants to make memory smart. The idea has a
small, determined following. Put everything on a single chip, say the devotees,
and you can make superfast minicomputers, perfect for intensive operations such
as searching the Library of Congress, or for hand-held devices that recognise
words of spoken English and spit them out in Portuguese. With nurturing, IRAM
could even become the brains behind the next generation of supercomputers.

Open up your desktop computer and you鈥檒l find its brain split in twain. The
logic, or microprocessor, lives on one chip; the memory on another. The two are
connected by a bundle of wires called a bus. The problem is that those clumsy
wires strain communication between the logic, which does all the thinking, and
the memory, which sits next door like an idiot with perfect recall. Every time
the microprocessor needs some information, it sends a request to memory via the
bus. The memory digs up the data and sends it back, again via the bus. But it
takes a lot of power to send electrical signals down those wires and the metal
pins that connect them to the chips. The bus is also excruciatingly slow. It
typically has only a few dozen wires, and each can carry only a single binary
bit at a time.

And things are getting worse. Memory capacity and microprocessor speed have
quadrupled every three years since the mid-1970s, but the bus hasn鈥檛 kept pace.
As designers boast 鈥渟creamingly fast鈥� processors, the gap yawns wider. 鈥淚t鈥檚
called the memory wall,鈥� says Scott Stiffler, an engineer with IBM. That wall,
says Patterson, is the single biggest obstacle to improving computer
performance.

The solution is to remove the bottleneck by putting memory and logic on the
same chip. Then, tiny metal pathways can be etched on the chip to make a tiny
thousand-lane highway between the memory and the logic. Even better, the
journey time falls dramatically. Data flow like water from a firehose, up to 10
trillion bits a second, a thousand times faster than today鈥檚 best bus. 鈥淚t鈥檚
pretty phenomenal,鈥� says Patterson.

In fact, designers have been edging this way for some time, sneaking larger
and larger chunks of fast memory, called static random access memory, onto their
microprocessors. But this SRAM (known as cache) takes up a lot of space so they
have to make do with just enough to serve as a kind of short-term memory. It
eliminates the need for data to be continually shuffled back and forth across
the bus. Patterson鈥檚 interest, however, is not in SRAM, but in Dynamic Random
Access Memory, which already forms the core memory of a PC.

Leaky memory

DRAM and SRAM (like microprocessors themselves) are just collections of
transistors. To store a binary bit in SRAM takes six transistors arranged to
make a kind of electronic seesaw that can be locked in one of two states, 1 or
0. By contrast, DRAM needs only a single transistor to charge up a tiny
capacitor. If the capacitor is charged up, it鈥檚 storing 1; if it鈥檚 empty it鈥檚 0.
With DRAM you can fit at least six times the data into the same space as you can
with SRAM. The problem is that capacitors leak their charge, so they have to be
tested, and any charge topped up, about 16 times a second. The higher the
temperature, the faster they leak, so to keep them cool the chips are designed
to consume low power. That鈥檚 good because they don鈥檛 drain battery power. The
disadvantage is that refreshing the capacitors slows DRAM down a lot.

Why doesn鈥檛 everyone already put DRAM on their processors? Until recently,
few companies had a factory that could put both on the same chip. Logic
transistors need to be fast, DRAM transistors need to be dense, and the two are
manufactured very differently. But sacrifice a bit of density and you can make
transistors good for both. Even so, the move towards hybrid chips 鈥渋s going to
be an evolution, not a revolution鈥�, says Betty Prince, a Texas-based industry
analyst and author of numerous books on the DRAM market. There is, however, a
revolution in thought. 鈥淒RAM is like a Rorschach test,鈥� Patterson says. People
see all kinds of possibilities.

Patterson鈥檚 vision is to revive vector processing, an idea once used in Cray
super-computers. This makes for a simple-minded computer, but one that excels at
doing dozens of identical things simultaneously. Vector processors don鈥檛 just
multiply two numbers together, they鈥檒l do the same for two long columns of
numbers without blinking. Such vast computational demands are common in
multimedia applications. The chips can easily be programmed, Patterson says, to
deal with the quick shifts of perspective that happen when you 鈥渨alk鈥� through a
virtual building, say. In recent years, vector processing has fallen into
disrepute. 鈥淚t鈥檚 kind of retro,鈥� says Patterson, a 1970s-style computer built
with 1990s technology.

Unexplored territory

Vector processors are simple enough that Patterson can design them with a
handful of graduate students. That in itself is retro. Twenty years ago, the
field was so young that a gang of caffeinated graduate students could still
shock industry giants with something they had built in a basement. But today鈥檚
microprocessors have so many complex parts that designing one is like trying to
construct New York from scratch all at once. The fabrication plants that carve
brains out of silicon come with billion-dollar price tags, and only a few
huge companies can play the game.

But intelligent memory is largely unexplored territory and the little guys
are back in the game, each with their own interpretation of the inkblot. When
Jack Lipovski, a computer scientist from the University of Texas at Austin,
looks at combined logic and DRAM, he sees a search engine. Databases, e-mail
archives and websites, he points out, are growing like weeds. 鈥淲e have all this
damn data out there and no one can search it!鈥� he cries. Lipovski thinks the
hybrid chip could be just the power tool for hacking through the
undergrowth.

Instead of putting one central processor on a chip, as Patterson proposes,
Lipovski wants to install scads of little ones, each capable of handling a bit
of arithmetic. On a DRAM chip, data are stored in a huge grid like a wall of
postboxes. Each bit lives in a separate box with a unique row and column
address. Typically, to search for a certain sequence of bits, the contents of
each box has to be retrieved and studied by the CPU until it finds a match.
That鈥檚 slow. But place an army of simple processors at the top of the columns,
Lipovski says, and every column can be searched at the same time, speeding up
the process dramatically. Lipovski calls the idea Dynamic Associative Access
Memory.

Who needs these DAAM chips? The human genome project for a start. One
person鈥檚 DNA, Lipovski says, will fill about 2 gigabytes of memory. To search
this from start to finish for a loosely specified DNA sequence could take all
day. True, there are algorithms to speed it up, but they鈥檙e not perfect,
Lipovski says. But put that genome onto a DAAM chip so that each column
contained a section of the total, and you could finish the search in less than a
minute, he claims.

Lipovski thinks the chips could become ubiquitous. Hard drives, for example,
have about 20 magnetic heads that read different parts of the disc, but only one
is active at a time. Wire up a DAAM chip to the heads, and all 20 could run at
the same time. 鈥淵ou can search a 10-gigabyte disc in about 30 seconds,鈥� he says.
That could be great for finding, say, one book among the 100 million in the
Library of Congress.

A DAAM chip would also make a pretty good calculator, especially for huge
numbers. Adding two 100-digit numbers is faster, for instance, if you split the
task: you add the low digits and a friend adds the high digits, then combine
your results. Those little processors perched on top of the memory columns can
also tackle different parts of large mathematical operations. That鈥檚 useful for
scientists, but huge numbers also form the basis for the encryption that keeps
computer communications safe from prying eyes. The chips could be a fast way to
do the mathematical mangling that encodes the data. Or, with enough of them
running at the same time, they could crack those codes by trying all the
possible keys.

Lipovski鈥檚 chips aren鈥檛 ready yet, though he鈥檚 started a company called
Linden Technology in Austin, Texas, to develop them. For the time being, he鈥檚
got patents. 鈥淲e鈥檙e bristling with lawyers,鈥� says an enthusiastic Lewis Larson,
co-founder of the company. Larson says they expect Sanyo to put the first DAAM
chips together in the next few months.

Hybrid chips, however, have already made a few people rich. Since 1995,
NeoMagic Corporation of Santa Clara, California, has been making a chip to
handle graphics for laptops that is part-DRAM, part-logic. Laptop screens have
to be refreshed about 60 times a second. Data for the image used to be stored on
a memory chip, driven across a bus to a logic chip which would prepare it for
display, then sent to the screen鈥攁 vast amount of traffic. NeoMagic sent
the bus to the scrap heap and relocated the whole operation onto a single
silicon wafer. 鈥淲e have the fastest chips in the market,鈥� says Neomagic鈥檚
president Prakash Agarwal. And without the bus sucking up power, the new chip
extends battery life, he says. The chips come standard in most new laptops.

The potential benefits have not been lost on companies that make memory,
either. Siemens, the largest maker of DRAM in Europe, is testing a hand-held
speech recognition unit that uses a combined DRAM and logic chip. 鈥淚t鈥檚 not on
the market yet,鈥� says Siemens鈥檚 marketing director in the US Gil Russel, 鈥渂ut it
looks pretty darn good.鈥� The chip digitises speech and searches for distinctive
scraps of sound called phonemes which it compares with templates stored in its
memory to find close matches. Even better is a prototype that translates speech
from one language to another. It鈥檚 crude at present, but Russel is undaunted.
鈥淚t鈥檚 absolutely stupendous,鈥� he gushes.

Such uses may seem small beer next to the versatile desktop computer. But
like the tiny microbes that make up most of the Earth鈥檚 biomass, it鈥檚 the small
unseen processors that make up most of the market. 鈥淓mbedded processors鈥�, which
fit inside everything from coffee-makers to fax machines, outsell personal
computers by almost 100 to 1. For makers of hybrid processors, then, this hidden
market seems like the sensible starting place.

Sure enough, Mitsubishi is making the first hybrid microprocessors in its
factory in Saijo, Japan. 鈥淲e鈥檝e been designing them behind closed doors for at
least four or five years,鈥� says Eric Nguyen, a manager at Mitsubishi. The M32RD
with its small processor is the first real 鈥渃omputer on a chip鈥�, Nguyen says. He
reckons that with an Internet connection and a keyboard, the M32RD is perfect
for a low-cost gadget to turn a TV into a window on the World Wide Web. The chip
has enough memory and logic to run a Web browser. Mitsubishi also hopes to
install the M32RD in everything from mobile phones to washing machines. The cost
per chip is about $20鈥攁 Pentium II costs several hundred
dollars.

Mere mortals

Whether merged logic and DRAM will ever replace microprocessors inside
personal computers, no one knows. Fred Pollack, a researcher at Intel, tested
the idea and found that for applications such as word processors, the hybrid
chip was slower. 鈥淚t鈥檚 only good for scientific math,鈥� Pollack says, 鈥渘othing
that would be used by mere mortals.鈥� Patterson says that鈥檚 because computers
today aren鈥檛 set up to exploit the easy access to data that hybrid chips would
offer.

At the University of Notre Dame, Indiana, computer scientist Peter Kogge is
designing a computer from scratch to do just that. Kogge started building hybrid
chips when he was at IBM in the 1980s and got hooked on the idea. Today, he
works with half a dozen universities and national laboratories on a project
called Hybrid Technology MultiThreading鈥攁n attempt to build a
supercomputer that runs a thousand times faster than anything around today. If
the machine comes together, it could replace today鈥檚 supercomputers in modelling
things such as nuclear explosions, supernovae or the climate.

But even if tomorrow鈥檚 supercomputers use hybrid chips, Kogge thinks it could
be a while before the technology is adopted for PCs. Most software around today
is written for IBM or Macintosh-style computers, and switching everyone to a
completely new machine will be an uphill task. Nobody wants to rewrite those
programs to run on a new kind of machine. Evolution doesn鈥檛 necessarily favour
the fastest machine, says Kogge, but the one that survives in the
marketplace.