The researchers are wrestling with ways to make a computer carry out 1000
trillion floating point operations per second. Today’s supercomputers can only
manage a trillion operations per second.

The scientists think the key is what they call a hybrid multi-threaded (HMT)
design. The computer would have 4000 parallel superconducting processors
supplied with data by superconducting memory chips. Long-term memory would use
conventional semiconductors.

While superconducting processors do not exist yet the team does know that the
HMT’s speed creates problems. The clock speeds of the chips should be at least
100 gigahertz—around 500 times the speed of the Pentium processors of PCs
now in the shops. But the fastest semiconductor memory could not run at more
than about a quarter of a gigahertz. This would force a processor to spend lots
of time waiting for data from memory.

The multi-threaded design will help to overcome this problem. A Pentium
processor is single-threaded—having only one queue of operations. A
multi-threaded machine has as many as 128 queues at every processor, and each
one is dealt with in turn.

Even so, fetching data from some parts of this computer’s memory could take
up to 1 million clock cycles. So even with 128 threads, the processor would be
waiting a long time for relevant data.

To solve the problem the team has come up with preparation units. These would
examine the instructions at the front of each queue, determine what data the
processor will need to execute them, and put the necessary data into
superconducting memory near the processor. This approach could mean that as few
as 64 threads will be enough to offset the disparity in speed between the
processors and the memory.

Despite this early work, the researchers say it will be at least ten years
before an HMT computer is built.

The tiny semiconductor switch uses a ring of gallium arsenide just 10.5
micrometres in diameter. Deana Rafizadeh, the graduate student who made the
switch, says it could be even smaller—between 1 and 2 micrometres across.
Previously, the smallest ring-shaped optical resonator was a mighty two
millimetres across.

Rafizadeh made her switch, which she describes as a nanoscale waveguide-coupled
microcavity resonator, at Cornell University’s Nanofabrication
Laboratory. The finished device consists of the resonator ring and two straight
waveguides just 0.5 micrometres wide that pass within 0.1 micrometres of
opposite sides of the ring. The waveguides channel light into the resonator. If
the ring is resonating at the same frequency as the light, the light passes
across. If not, it stops. Rafizadeh says, “If the light is in resonance with the
ring, then the incoming light is fully coupled from one waveguide into the ring
and out the other waveguide.”

The ring can be tuned by applying a voltage or changing its temperature. This
alters its refractive index, and hence its resonant frequency. Once tuned, this
resonator acts like an extremely selective filter that allows only light within
a very narrow band of frequencies to pass.

The resonator can be made to separate many more frequencies than larger
devices can. Such a tunable resonator could perform a variety of tasks,
including filtering one frequency of light from another and combining, or
switching between, many different communication lines.

Despite this success, the researchers are still a long way from producing an
integrated circuit that runs on light alone. The mechanics of controlling the
resonator have to be fully worked out. What’s more, the resonator only works
with photonic-wire lasers built to the same scale. The first such laser was
developed in 1995, by the same laboratory at Northwestern University. Seng-Tiong
Ho, head of the laboratory, says the first completely optical integrated circuit
lies a couple years away.