Image: Christine Daniloff |
Computer
chips have stopped getting faster. In order to keep increasing chips’
computational power at the rate to which we’ve grown accustomed, chipmakers are
instead giving them additional “cores,” or processing units.
Today,
a typical chip might have six or eight cores, all communicating with each other
over a single bundle of wires, called a bus. With a bus, however, only one pair
of cores can talk at a time, which would be a serious limitation in chips with
hundreds or even thousands of cores, which many electrical engineers envision
as the future of computing.
Li-Shiuan
Peh, an associate professor of electrical engineering and computer science at Massachusetts
Institute of Technology (MIT), wants cores to communicate the same way computers
hooked to the Internet do: By bundling the information they transmit into “packets.” Each core would have its own router, which could send a packet down
any of several paths, depending on the condition of the network as a whole.
At
the Design Automation Conference, Peh and her colleagues will present a paper
she describes as “summarizing 10 years of research” on such “networks on chip.”
Not only do the researchers establish theoretical limits on the efficiency of
packet-switched on-chip communication networks, but they also present
measurements performed on a test chip in which they came very close to reaching
several of those limits.
Last stop for buses
In principle, multicore chips are faster than single-core chips because they
can split up computational tasks and run them on several cores at once. Cores
working on the same task will occasionally need to share data, but until
recently, the core count on commercial chips has been low enough that a single
bus has been able to handle the extra communication load. That’s already
changing, however: “Buses have hit a limit,” Peh says. “They typically scale to
about eight cores.” The 10-core chips found in high-end servers frequently add
a second bus, but that approach won’t work for chips with hundreds of cores.
For
one thing, Peh says, “buses take up a lot of power, because they are trying to
drive long wires to eight or 10 cores at the same time.” In the type of network
Peh is proposing, on the other hand, each core communicates only with the four
cores nearest it. “Here, you’re driving short segments of wires, so that allows
you to go lower in voltage,” she explains.
In
an on-chip network, however, a packet of data traveling from one core to
another has to stop at every router in between. Moreover, if two packets arrive
at a router at the same time, one of them has to be stored in memory while the
router handles the other. Many engineers, Peh says, worry that these added
requirements will introduce enough delays and computational complexity to
offset the advantages of packet switching. “The biggest problem, I think, is
that in industry right now, people don’t know how to build these networks,
because it has been buses for decades,” Peh says.
Forward thinking
Peh and her colleagues have developed two techniques to address these concerns.
One is something they call “virtual bypassing.” In the Internet, when a packet
arrives at a router, the router inspects its addressing information before
deciding which path to send it down. With virtual bypassing, however, each
router sends an advance signal to the next, so that it can preset its switch,
speeding the packet on with no additional computation. In her group’s test
chips, Peh says, virtual bypassing allowed a very close approach to the maximum
data-transmission rates predicted by theoretical analysis.
The
other technique is something called low-swing signaling. Digital data consists
of ones and zeroes, which are transmitted over communications channels as high
and low voltages. Sunghyun Park, a PhD student advised by both Peh and Anantha
Chandrakasan, the Joseph F. and Nancy P. Keithley Professor of Electrical
Engineering, developed a circuit that reduces the swing between the high and
low voltages from one volt to 300 millivolts. With its combination of virtual
bypassing and low-swing signaling, the researchers’ test chip consumed 38% less
energy than previous packet-switched test chips. The researchers have more work
to do, Peh says, before their test chip’s power consumption gets as close to
the theoretical limit as its data transmission rate does. But, she adds, “if we
compare it against a bus, we get orders-of-magnitude savings.”
Luca
Carloni, an associate professor of computer science at Columbia University who
also researches networks on chip, says “the jury is always still out” on the
future of chip design, but that “the advantages of packet-switched networks on
chip seem compelling.” He emphasizes that those advantages include not only the
operational efficiency of the chips themselves, but also “a level of regularity
and productivity at design time that is very important.” And within the field,
he adds, “the contributions of Li-Shiuan are foundational.”
