IEEE Transactions on Computers, Journal Year: 2023, Volume and Issue: 73(2), P. 366 - 379
Published: Nov. 7, 2023
The
rapid
development
of
deep
learning
has
propelled
many
real-world
artificial
intelligence
applications.
Many
these
applications
integrate
multiple
neural
networks
(multi-NN)
to
cater
various
functionalities.
There
are
two
challenges
multi-NN
acceleration:
(1)
competition
for
shared
resources
becomes
a
bottleneck,
and
(2)
heterogeneous
workloads
exhibit
remarkably
different
computing-memory
characteristics
synchronization
requirements.
Therefore,
resource
isolation
fine-grained
allocation
each
task
fundamental
requirements
computing
systems.
Although
number
acceleration
technologies
have
been
explored,
few
can
completely
fulfill
both
requirements,
especially
mobile
scenarios.
This
paper
reports
Hierarchical
Asynchronous
Parallel
Model
(HASP)
enhance
performance
meet
HASP
be
implemented
on
multicore
processor
that
adopts
Multiple
Instruction
Data
(MIMD)
or
Single
Thread
(SIMT)
architectures,
with
minor
adaptive
modification
needed.
Further,
prototype
chip
is
developed
validate
the
hardware
effectiveness
this
design.
A
corresponding
mapping
strategy
also
developed,
allowing
proposed
architecture
simultaneously
promote
utilization
throughput.
With
same
workload,
demonstrates
3.62
Language: Английский