15 May 2018
| Alan Burkitt-Gray
Google and NEC have doubled the performance of the transpacific Faster cable linking Taiwan to Oregon, and believe they can triple it.
The 10,000km cable was originally designed to carry traffic
at 2 bits/second/hertz, say the companies, but using artificial
intelligence (AI) techniques they increased spectral efficiency
first to 3b/s/Hz and then to 4b/s/Hz – twice the
original design specification.
Toru Kawauchi, general manager of NEC’s submarine
network division, said: "This approach sets aside those
[previous] deterministic models of nonlinear propagation, in
favour of a low-complexity black-box model of the fibre,
generated by machine learning algorithms."
Google and NEC tested the performance in a field trial
performed together with live traffic neighbouring channels.
They claim a spectral efficiency-distance product record of
66,102b/s/Hz. It the trial it carried live traffic from Google
NEC has also carried out offline field trials over dark fibres
with even more promising results. They used fibre of the same
length and achieved spectral efficiency of 5.68b/s/Hz and made
other measurements that promise 6.06b/s/Hz – three
times the original design specification of the Faster cable.
NEC said it believes 6b/s/Hz is a "realistic target".
Kawauchi said: "The results demonstrate both an improvement in
transmission performance and a reduction in implementation
complexity. Furthermore, since the black-box model is built up
from live transmission data, it does not require advance
knowledge of the cable parameters."
The team achieved this result using a technique called
probabilistic-shaping which achieved a performance close to the
theoretical maximum as predicted by Claude Shannon, the Nobel
prize-winning Bell Labs mathematician who is known as the
father of information theory for his work in the 1940s and
"For the first time on a live cable, AI was used to analyse
data for the purpose of nonlinearity compensation," said NEC in
"Other approaches to NLC have attempted to solve the nonlinear
Schrödinger equation, which requires the use of very
complex algorithms," said Kawauchi. This is a technique named
after another Nobel prize-winner, Erwin Schrödinger.
Kawauchi added: "This allows the model to be used on any cable
without prior modelling or characterisation, which shows the
potential application of AI technology to open subsea cable
systems, on which terminal equipment from multiple vendors may
be readily installed."
The companies published their results in a scientific paper (PDF) authored by three
Google scientists and six from NEC.