Transcript HiPEAC view on NMC

High-Performance and
Embedded Architecture and
Compilation
HiPEAC Vision
2017
on
Neuromorphic computing
Marc Duranton
Koen De Bosschere, Christian Gamrat,
Jonas Maebe, Harm Munk, Olivier Zendra
The HiPEAC project has received funding from the European Union’s Horizon 2020
research and innovation programme under grant agreement number 687698.
(Narrow) Artificial Intelligence
everywhere
• Artificial Intelligence (and Deep Learning) is
changing the man-machine interaction –
natural interfaces, ”intelligent” behavior
• The new systems should make intelligent and
trustable decisions (that can be also explained)
2
Human and machine collaborating
• Entering the Centaur1 era
• Intelligent Personal Assistant
(Siri, Cortana, Google now, Alexa…)
• Self-Driving car
• BIC (Brain Inspired Computing)
• …Mainly using Deep Learning
1 In Advanced Chess, a "Centaur" is a man/machine team.
techniques for natural signal processing
Advanced Chess (sometimes called cyborg chess or centaur
chess) was first introduced by grandmaster Garry Kasparov,
with the objective of a human player and a computer chess
program playing as a team against other such pairs.
3
(from Wikipedia)
(Deep) Learning is quite demanding
Example of hardware: Baidu’s Minwa
– For vision using deep learning
– 36 server nodes, each with Intel Xeon E5-2620,
FDR Infiniband (56Gb/s) and 4 nVisia Tesla K40m
GPU
– Total of 8.6 TB of memory
Example of hardware: NVIDIA DGX-1
Customized hardware…
… required to increase energy efficiency
(for the inference phase)
Computations (operations and precision) adapted to the use
6
AI and Deep
Learning
techniques
are not only
lab research,
but are
already at
customer
level
7
Source: 2017 Accenture CMT Digital Consumer Survey
Interoperability and composability
Multiple Control Apps
Interoperability and composability solutions are required
But why voice -> text and text -> voice run on the Cloud?
8
Computing Distribution for ”Cognitive” systems
Real-time
Embedded
Intelligence
at the edge
Sensors
(Video, Sound, …)
Cyber Physical
Physical Systems
Entanglement
New
services
Cloud / HPC
“Dumb” Internet of
Things devices
Processing,
Abstracting
Understanding
as early as
possible
(often only
inference
required for
Deep Neural
networks)
Big Data
Data Analytics /
Cognitive computing
/ Deep learning
Transforming data into information
as early as possible
9
Embedded intelligence needs
local high-end computing
System should be autonomous to
make good decisions in all
conditions
Safety will impose that basic
autonomous functions
should not rely on “always
connected” or “always
available”
Cloud and HPC cannot support many
cyber-physical applications.
10
Embedded intelligence needs
local high-end computing
Example: detecting elderly
people falling in their home
Privacy will impose that some processing
should be done locally and not be sent to the cloud.
11
Embedded intelligence needs
local high-end computing
Dumb sensors
Smart sensors: image and
signal (voice) recognition
Bandwidth will require more local processing
12
Managing complexity
Cognitive solutions for
computing systems:
• Using AI techniques
for computing systems
• Similar to Generative
design for mechanical
engineering
13
Generative design approach
The user only states desired goals and constraints
Motorcycle swingarm: the piece that hinges the rear wheel to the bike’s frame
14
Highlights of HiPEAC vision 2017…
Cyber physical entanglement
Artificial intelligence
Human and machine collaboration
15
Download the
new HiPEAC Vision at:
http://hipeac.net/vision
Give us your
comments at:
[email protected]
16
Further (short) reading
Read the EETimes interview at:
http://www.eetimes.com/document.asp?doc_id=1331
052&page_number=1
17
Evolution of man
Yesterday
Today
Tomorrow
18
•
NVM Synapses Implementations
and spiking neural networks
2-PCM synapses for unsupervised cars trajectories extraction
PCM
From spiking pre-synaptic
neurons (inputs)
VRD
ILTP
ILTD
Crystallization/
Amorphization
I = ILTP - ILTD
Spiking postsynaptic neuron
Equivalent
(output)
2-PCM synapse
[O. Bichler et al., Electron Devices, IEEE Transactions on, 2012]
•
CBRAM binary synapses for unsupervised MNIST handwritten digits classification
with stochastic learning
CBRAM
Forming/Dissolution of
conductive filament
[M. Suri et al., IEDM, 2012]