IEEE Solid-State Circuits Magazine - Fall 2015 - 60

As the distance to the chip becomes greater,
the required switching energy of the data-transfer
link becomes larger.
hardware problems are obtained by
using devices other than transistors.
Neural-network inference is highly
efficient partly because the operation is
programmed as the connection weight
of the artificial neurons, with the output
obtained when the data passes through
the resulting network (Figure 10).
Because the connection weight is ideally kept locally close to the inferring
neuron, it is not necessary to obtain
instructions by accessing a memory or
by temporarily storing transactions in
a memory. This kind of programming,
which uses a connection network, is
commonly found in FPGAs and other

devices targeting higher efficiency,
such as quantum annealing [13].

The Future of Wireline Technology
Now, let us forecast the role of chipto-chip wireline communication to
achieve the performance improvements summarized in the previous section. Assuming the rate of
increase in power efficiency to be
tenfold every five years and allocating this to energy reduction rates for
compute operation, memory access,
and data transfer, the demands to
be placed on wireline technology
can be derived.

Relative Energy
Relative Ferquency

10-1

10-2

Total
Processing
Move
Move Frequency

10-3

10-4
2015

2020

2030

2025
Time (Year)

Figure 11: processing energy and data-move energy.

TablE 3. NORMalIzEd ENERgIES Of a COMpuTER SySTEM.
2015

2020

2025

Total energy

0.1

0.01

Processing

0.01

Data move
B/F ratio
Transceiver
efficiency

fa l l 2 0 15

2030

0.001

To perform an information-processing task, compute operations
and data transfers should take place.
Let us assume that the sum of the
compute operation energy per task
and the data-transfer energy is normalized by its value in 2015. Then,
the target values for the total energy
after five, ten, and 15 years will be
10 -1, 10 -2, and 10 -3, respectively,
relative to the value in 2015.
Compute-operation energy and
data-transfer energy are allocated
as follows. As of 2015, the operation energy is assumed to be 1%
of the total energy of the system.
Operation energy shall be determined by assuming that scaling will
continue until 2020. After 2020, no
further scaling can be expected,
and compute-operation energy will
be reduced at a certain rate only by
improving the associated computing
algorithms. By 2030, the energies
for data transfer and for operation
will be equal at 0.5 # 10 -3 . The reason for this is that by 2030, where
energy must be reduced to the limit,
data transfer and operation will be
conducted in an inseparable manner, leaving no room for one to be
significantly greater than the other.
Once the operation energy has
been determined, data-transfer energy
is obtained by subtracting the operation energy from the total energy.
The resulting trend in the switching
energy of wireline communication
transceivers has been forecast assuming that in 2030 the use of improved
computing schemes will cause the
frequency of data-transfer events per
information-processing task to reduce
exponentially-by a factor of 10 -2
compared to its value in 2020.
The future of wireline communications technology obtained from
the assumptions discussed earlier is
described in the following and summarized in Figure 11 and Table 3.

0.004

1.4 # 10

-3

0.5 # 10 -3

0.99

0.096

8.6 # 10 -3

0.5 # 10 -3

2015-2020

0.4

0.06

0.01

Normalized

0.24

0.15

0.05

Absolute [pJ/b]

0.96

0.6

0.2

In this period, I assume that operation
energy is proportional to the scaling
factor x. The current technology is
assumed to be somewhere between

IEEE SOLID-STATE CIRCUITS MAGAZINE

Table of Contents for the Digital Edition of IEEE Solid-State Circuits Magazine - Fall 2015