science professor David A. Patterson gazes into his crystal ball
and lays out the future, the industry listens.
And what they're hearing now is the sound of an IBM chip fabrication
plant cranking out the first full prototype of Patterson's IRAM
(Intelligent Random Access Memory), a single chip that combines
a microprocessor with memory to increase a computer's speed while
reducing its hunger for power. If it works as expected, IRAM
with its 100 million transistors jammed onto a single 18.5 mm
square bit of silicon - will be the ideal brain for next-generation
mobile phones and handheld personal computers.
"IRAM makes for a performance bonanza," says Patterson,
who developed the simple yet profound concept with Berkeley computer
science professor Katherine Yelick and a small team of dedicated
graduate students. "And, it has the advantage of being smaller
and more energy-efficient than current architectures."
Patterson is known for predicting the future by inventing it.
Two decades ago, he raved about a new computer architecture called
RISC reduced instruction-set computing an against-the-grain
computer architecture he and his students developed to simplify
computer chips by dolling out an increased number of tasks to
the software. When the speedy and low-cost RISC chips made their
commercial debut in the mid-1980s, the microprocessor market was
changed forever. After igniting the RISC revolution, Patterson,
in collaboration with professor Randy Katz, focused his forecasts
on disk drives. The result was RAID (Redundant Arrays of Inexpensive
Disks), clusters of fast and cheap disk drives working in tandem.
Introduced in 1986, RAID is now the core of a thirty billion dollar
per year file server industry.
Following in that illustrious lineage, IRAM has the potential
to become the architecture-of-choice in tomorrow's post-PC marketplace
where information technology is everywhere, not just on the desktop.
The researchers envision personal digital assistants (PDAs) that
take dictation, cellular phones with perfect speech recognition
and video features, and TV set-top boxes that deliver incredibly
fast and lifelike interactive graphics to your living room.
her student years at MIT, Katherine Yelick (right) rowed
on the undergraduate and graduate crew teams. These days
though, her passion is playing with her two toddlers, ages
three and five, and a bit of biking, backpacking, and running.
(Click for larger image.)
To understand what makes IRAM so smart, you have to know what's
under the hood of almost every computer. Start with dynamic random
access memory (DRAM) chips, the scratch pad for software. Data
is temporarily stored in the DRAM for quick access by the computer's
brain, the microprocessor chip. The problem is that the two components
cant communicate quickly enough to take full advantage of
the blazing speeds of today's microprocessors. (Today's microprocessor
are 100 times faster than their 20-year-old ancestors while memory
chips have only seen a tenfold speed increase.) The result is
latency, time wasted by the microprocessor waiting for the data
it needs from DRAM. The other problem faced by today's microprocessors
is available bandwidth, the amount of data that can be transmitted
in a given amount of time. To keep cost and power consumption
low - critical when it comes to portable electronics - today's
DRAM designs are limited to just a handful of connections to the
With IRAM though, data can fly between the memory and the microprocessor
through a multitude of short pipes that don't consume nearly the
power of chip-to-chip hops. Meanwhile, latency is improved simply
because the time a signal takes to travel between the processor
and the memory is minimized by the components' proximity. An added
bonus is that IRAM's memory and processor are built from just
two distinct kinds of modular building blocks that can easily
be increased or decreased in number depending on the performance
and price goals of the manufacturer.
But according to Berkeley's IRAM researchers, the real power
behind the IRAM architecture comes from a fundamental change in
the way the processor does its job. To bring out the best in IRAM,
Yelick revisited a 30-year-old concept called vector processing,
invented to perform scientific calculations on room-size mainframe
computers. Based on code donated from supercomputer superpower
Cray Inc., Yelick developed software that enables IRAM's vector
processor to be programmed in todays common computer tongues.
Ultimately though, whether IRAM ends up inside tomorrow's mobile
phones, PDAs, and interactive televisions is up to the industry.
"What we've always done was to prove an idea with a prototype
so that commercial companies could build newer and bigger things
by putting more than a half-dozen people on the project," Patterson
says. "Our goal really is just to convince skeptics and inspire
others to take our work further."
The Berkeley IRAM Project
David A. Patterson's