IEEE Solid-States Circuits Magazine - Fall 2023 - 24

Figure 4 shows the number of
NPUs in different types, as published
in the proceedings of the International
Solid-State Circuits Conference (ISSCC),
IEEE Journal of Solid-State Circuits
(JSSC), IEEE Transactions on Circuits
and Systems (TCAS) I, TCAS II, IEEE
Transactions on Biomedical Circuits
and Systems, and IEEE Transactions
on Very Large Scale Integration (TVLSI),
with steps of two years. It can be
seen that type II is the main stream of
the NPUs in brain-machine interfaces
(BMIs) during the past decade. Type I is
gradually eliminated in the top publications.
Since 2019, type III has started
to show its advantages.
Before ML was widely used in hardware,
the standard method for classification
was to set a threshold for an
extracted feature, which was applied
to all the type I NPUs. To perform lowpower
on-chip processing, the following
trends of type I NPUs have been
witnessed:
1) the use of sub-Nyquist processing
such as compressive sensing [37]
2) The solution of multiplier free
processing [36]
3) The use of analog computing [37]
4) The use of more flexible classifiers
[38].
Compared with the processors in
type I, the threshold is replaced by
several more complex ML classifiers
including SVM, decision tree
(DT), and linear least square classifiers.
It could be highlighted that the
combination of the spectral band en5
6
7
1
2
3
4
Year
FIGURE
4: The number of NPUs in different types.
24
FALL 2023
IEEE SOLID-STATE CIRCUITS MAGAZINE
Type 1
Type 2
Type 3
ergy feature with an SVM classifier
seems the most popular selection. In
type II NPUs, we highlight the trends
of general-purpose NPUs and onchip
learning:
1) the trend of general-purpose NPUs
[41], [44], [46]
2) the trend of on-chip learning
[41], [46], [48], [49], [50]
3) toward the NN [42].
The NNs have shown their ability
in several challengeable tasks. The
classification metrics of several biosignal
processing tasks are still unsatisfactory,
which produces a crucial
requirement for the hardware implementation
of NNs. In the type III
NPUs, the following two trends are
highlighted.
1) the trend of low-complexity computing
[52], [54]
2) the trend of end-to-end NNs [56].
There is always a tradeoff between
the accuracy and battery life of NPUs.
Due to the unsatisfied accuracy, the
type I NPUs have been eliminated in
the past years. To achieve higher accuracy,
NPUs are in the process of upgrading
from traditional ML based to
deep learning based in recent years.
During the development of each type
of NPUs, multiple technologies can
be seen that aim to reduce power
consumption in order to increase
the battery life, such as sub-Nyquist
processing, multiplier free processing,
analog computing, and low-complexity
computing. From the existing
research results, the end-to-end NNbased
NPU with optimized low computational
load is promising to be a
competitive solution in the future.
Low-Power Wireless Transceiver
A reliable wireless data link is important
to support free movement for
animal subjects. However, the wireless
module becomes a bottleneck to
increase the battery life of the neural
interface while the channel number
increases. Tradeoffs must be taken
into consideration, including the
data rate, power consumption, transmission
distance, system volume,
regional/worldwide medical regulations,
and any specific requirements
pertinent to targeting applications.
Two directions of wireless transmission
are included in the neural
interface applications. The data rate
of the uplink including the collected
neural signals can be several Mb/s
for EEG, ECoG, and LFP signals, while
the data rate of the downlink is far
less since only control commands
are needed. As a result, the uplink
data transmission is more essential.
The narrow RF bands allocated
for medical devices in most countries
adhere to the standards of the
MICS, Medical Device Radiocommunications
Service, or ISM radio bands.
These regulations provide spectrum
resources including ~400 MHz,
~900 MHz, and ~2.4 GHz, offering
available bandwidth from 100 kHz to
several MHz. When adopting simple
modulation schemes, such as OOK,
BPSK, and FSK, the achievable data
rate with such bandwidth ranges from
several kb/s to several Mb/s [57], [58],
[59], [60], [61]. Thus, the transmission
of the acquired tens of channels
of LFP signals can be covered. Stateof-the-art
transmitters with a 0-dBm
output power supporting a transmission
range of several meters usually
feature an energy efficiency of 2 to
7 nJ/bit. When there is a demand for a
higher data rate, more complex modulation
methods, such as QAM, can
be employed. While this effectively
increases the data rate, the downside
is a more complicated transmitter
circuit design and an increase
2011-2012
2013-2014
2015-2016
2017-2018
2019-2020
2021-2022
Chip Number
IEEE Solid-States Circuits Magazine - Fall 2023

Table of Contents for the Digital Edition of IEEE Solid-States Circuits Magazine - Fall 2023
