Since that time, companies, particularly NVIDIA, have helped lead the way
toward an artificial intelligence revolution with the use of GPUs [39]. GPUs are
now utilized extensively by companies and universities involved in the fields of big
data, machine learning, and genomics, among others [39–41]. GPUs are also seeing
increased utilization in areas of academic research as more coding languages and
libraries, such as C and FORTRAN, are updated to take advantage of the parallel
processing power of this architecture.
3.4 Purpose-Built Chips and History
With an exponentially growing interest in neuromorphic architectures, many
researchers began further pursuing their development. In 2005, Fast Analog Computing with Emergent Transient States (FACETS) launched a research initiative
funded by the UK to implement brain-like hardware architecture for neuromorphic
computing. The project concluded in 2010 with a VLSI implementation of a
traditional CMOS fabricated device containing 400 neurons and 100,000 synapses
[42]. In 2011, brain-inspired multiscale computation in neuromorphic hybrid systems (BrainScaleS) intended to expand upon FACETS work and concluded in 2015
with an architecture containing 1.6 million neurons and 400 million synapses. In the
middle of the BrainScaleS project further interest in neuromorphic architectures
arose.
In 2013, the Information and Communication Technologies (ICT) of the EU
launched the Human Brain Project. This comprehensive research initiative aimed to
advance our understanding of the human brain through numerous fields including
neuroscience and computation. In the field of neuromorphic architectures, this
flagship intends to refine and expand upon the work of FACETS and BrainScaleS.
The same year, the United States launched the Brain Initiative under the Obama
administration with similar intent and currently funds numerous agencies including
DARPA, NIH and NSF. Through this funding neuromorphic based projects continue
to progress.
As these architectures advance, hardware has begun to be designed for task
specific applications. IBM has developed TrueNorth, a brain inspired device suitable
for complex applications which utilize neural networks. TrueNorth is a 5.4 billion
transistor chip containing 4096 neurosynaptic cores interconnected to an intrachip
that utilizes 1 million programmable spiking neurons and 256 million configurable
synapses; it consumes a mere 70 mW and is able to process 46 billion SOPS, per
watt. This greatly exceeds the limit of energy efficient super computers which are
only capable of processing 4.5 billion FLOPS per watt. This device has demonstrated high fidelity multi-object recognition in real-time capable of discerning
objects based on different classes (i.e. person vs. cyclist). TrueNorth highly excels
in task specific applications which do not require reconfigurability similar to ASICs.
The Spiking Neural Network Architecture (SpiNNaker) is on ongoing project
which aims to have 1 million ARMs processors in parallel. Unlike TrueNorth which
212
R. Aguilera et al.
toward an artificial intelligence revolution with the use of GPUs [39]. GPUs are
now utilized extensively by companies and universities involved in the fields of big
data, machine learning, and genomics, among others [39–41]. GPUs are also seeing
increased utilization in areas of academic research as more coding languages and
libraries, such as C and FORTRAN, are updated to take advantage of the parallel
processing power of this architecture.
3.4 Purpose-Built Chips and History
With an exponentially growing interest in neuromorphic architectures, many
researchers began further pursuing their development. In 2005, Fast Analog Computing with Emergent Transient States (FACETS) launched a research initiative
funded by the UK to implement brain-like hardware architecture for neuromorphic
computing. The project concluded in 2010 with a VLSI implementation of a
traditional CMOS fabricated device containing 400 neurons and 100,000 synapses
[42]. In 2011, brain-inspired multiscale computation in neuromorphic hybrid systems (BrainScaleS) intended to expand upon FACETS work and concluded in 2015
with an architecture containing 1.6 million neurons and 400 million synapses. In the
middle of the BrainScaleS project further interest in neuromorphic architectures
arose.
In 2013, the Information and Communication Technologies (ICT) of the EU
launched the Human Brain Project. This comprehensive research initiative aimed to
advance our understanding of the human brain through numerous fields including
neuroscience and computation. In the field of neuromorphic architectures, this
flagship intends to refine and expand upon the work of FACETS and BrainScaleS.
The same year, the United States launched the Brain Initiative under the Obama
administration with similar intent and currently funds numerous agencies including
DARPA, NIH and NSF. Through this funding neuromorphic based projects continue
to progress.
As these architectures advance, hardware has begun to be designed for task
specific applications. IBM has developed TrueNorth, a brain inspired device suitable
for complex applications which utilize neural networks. TrueNorth is a 5.4 billion
transistor chip containing 4096 neurosynaptic cores interconnected to an intrachip
that utilizes 1 million programmable spiking neurons and 256 million configurable
synapses; it consumes a mere 70 mW and is able to process 46 billion SOPS, per
watt. This greatly exceeds the limit of energy efficient super computers which are
only capable of processing 4.5 billion FLOPS per watt. This device has demonstrated high fidelity multi-object recognition in real-time capable of discerning
objects based on different classes (i.e. person vs. cyclist). TrueNorth highly excels
in task specific applications which do not require reconfigurability similar to ASICs.
The Spiking Neural Network Architecture (SpiNNaker) is on ongoing project
which aims to have 1 million ARMs processors in parallel. Unlike TrueNorth which
212
R. Aguilera et al.