By Simon Firth,
This is the first in a two-part series looking at
in photonics being pursued by researchers in HP’s Information and
Quantum Systems and Exascale Computing labs.
Photonics – the science of transmitting information via light – found
its first great commercial use in fiber optic cables spanning the
These optical cables formed the backbone of national telecommunications
networks, eventually carrying voice and data into homes and businesses.
Lately, they’ve been connecting networks of computers together within
those business locations.
“Now what we’re trying to do at HP Labs,” says researcher Mike Tan, “is
to break through to the next level to bring photonics into the computer
HP photonics researchers, from
left to right, Paul Rosenberg,
Sagi Mathai and Michael Tan
In particular, Tan explains, that means using light to
transmit data between the separate circuit boards, or blades, that
typically make up a modern server.
Today those connections are made by copper wires and pins. But as blade
circuitry gets ever faster, copper won’t be able to keep up.
Blade electronics currently run at 5 gigabits per second, and copper
can only go to 9.6 . Optical connections already run that fast, and in
the future will go to at least 20 gigabits per second.
Photonics, however, promises to do more than simply keep up with the
speed of blade circuitry.
“Not only will optical connections work faster,” explains Tan, “but
we’ve shown that they’ll use less power and they’ll provide new
connectivity between blades that’s not available today.”
Despite its potential, both technical and financial hurdles
have so far kept photonics outside the server box.
Cost is perhaps the biggest issue. Optical connections are expensive,
and each server requires hundreds.
“At the same time we're developing new concepts in connecting blades,”
says Tan, “we’re also using the telecommunications industry's existing
technology to try and do it in a very low-cost manner.”
One great advantage to building servers out of blades is that
each blade can be directly and independently plugged into the server’s
shared backbone, or mid-plane – a design that minimizes cable usage and
the labor required to connect those cables every time you change your
arrangement of blades.
Any switch from copper to optical connections would need the same ease
of connectivity, so HP photonics researchers first set their sights on
creating flexible optical connectors that would still allow blades to
be hot-swapped into and out of a server rack.
The researchers' innovation: a way to link two optical connectors
without having to connect them manually.
The secret lies in outputting the blade’s data via standard optical
converters -- vertical cavity surface emitting lasers (VCSELs) -- to
standard optical connectors that are attached to magnets. Clever
engineering ensures that when the magnets are anywhere near each other,
they first attract each other and then align their complex sets of
light beams perfectly.
These magnetically coupled interconnects not only get data off the
blade at much faster speeds than conventional copper connections, they
also allow the entire server backbone to be optically wired. That in
turn saves huge amounts of power, because electrical signals fade
quickly on copper lines, requiring frequent (power-hungry and
heat-generating) repeaters to boost their signal.
But what if you could eliminate the need for optical cables
altogether? You’d have the same increased speed and
power-efficiency, but gain an extraordinary level of flexibility in how
you connected the blades in a server.
A second major innovation from HP Labs – the Free Space Optical
Interconnect – does just that.
“We’re basically shooting optical beams across and between blades," Tan
Here's how it works: Researchers set up standard VCSELs so that they
slot in no more than 5cms apart from each other. The optical beams from
the transmitting and receiving VCSELs are then focused in such a way
that they’re automatically detected and aligned – allowing tolerance
for initial misalignment and adjusting for rotation and tilt of the
light beams using a set of actuators.
This design further reduces the need for signal repeaters and server
cooling. And because it makes possible direct optical connections
between individual blades, it also eliminates the need for the
traditional server backbone. That could create faster-running, and thus
more powerful, computers – even without increasing the processing power
of any individual blade.
“Essentially, you’ve enabled a more flexible computer architecture,”
suggests Tan, “because you now have connections which were previously
unavailable to you.”
High Speed Optical Multidrop Bus
A third recent innovation from HP photonics researchers uses
light to connect the circuitry of the blade itself.
As computer chips work at ever higher frequencies, the laws of
electronics make it harder to connect them in a parallel bus
configuration – because doing so requires ever more power, which adds
ever more capacitance, which in turn reduces transmission speed, which
means the components can’t ever connect as fast as they operate.
Because of this, the industry has turned to serial point-to-point
connectivity, which itself has problems with latency and the need for
ever more power to retransmit signals.
But HP’s new High Speed Optical Multidrop Bus allows a return to
parallel bus connectivity – even at very high speeds – by sending
multiple light beams along narrow hollow metal tubes called waveguides
between the components on a blade, and splitting or tapping a
small portion of the light so that it always has the same power
wherever it is in the circuit.
To keep costs low, researchers split-off the light with the
same techniques used to put coatings on a pair of eyeglasses.
benefits of the optical bus include making it possible to add more
memory to multi-core systems on a blade, without the power and latency
issues that come with serial point-to-point connectivity.
It’s not a bad set of innovations for a project that has been running
less than 15 months. But this is a team on the fast track.
“Within three to five years,” says Tan, “we want to have photonics
inside the rack.”