20
Hyesoon Kim Celebrating Yale@75 Sep 19 2014
Hyesoon Kim
Celebrating Yale@75 Sep 19 2014
2
3
| Energy, energy, energy or efficiency, efficiency, efficiency
| Many sensors
Camera (video)
Audio, Wi-Fi, GPS, temperature, health related sensors, accelerators
| Security, Privacy
vs.
| Cellular Neuron Network Processors
a neuron-network computing paradigm, used in variety of applications
Proposed by Chua and Yang in 80s.
| Very well suited for image processing applications
Only local connections short wire. low energy consumption
| Consume very little power
C(I,j)
Multip
lie
r
transducer
Processing Unit
Multip
lie
r
CNN cells
input image
output image
| Global information is passed through cells
| Other applications:
More complicated image processing algorithms
Optimal path finding
Character recognition
| Challenges of CNN
Image size scalability
Scalable CNN [GT]
Difficulty of programming
Learning templates
8
Em
bedded C
ellu
lar
Neura
l N
etw
ork
Multi-core CPUs
Special Accelerators
CLOUD
Video sensor processors are used for front-end processors Complex computations will be handled in multi-core CPUs
Multi-core
special accelerators
Memory
Memory
Memory
Memory
| Computation can be further offloaded to Cloud
| Performance and energy savings
| Q. What to offload?
| Q. Privacy, sensitive data
Key to have Wearable Computing
10
| Memory size
0
500
1000
1500
2000
2500
2006 2008 2010 2012 2014
Me
mo
ry S
ize
(M
B)
iphone iphone 3GS
Galaxy S iphone 4
Google Nexus 7
LG Optimus G
Galaxy S4
Galaxy S5
11
Current Mobile Systems Mobile system with NVM
12
0
2
4
6
8
10
12
14
Tim
e (
sec)
first launch resume
First launch takes several seconds.
| Dedicated Region for Start-up code
| Energy, energy, energy
| Many sensors
Camera (video)
Audio, Wi-Fi, GPS, temperature, health related sensors, accelerators
| Security, Privacy
vs.
| Tradeoff between security and performance
| Example: Google’s Native Client Platform (NaCl)
Run natively compiled applications in the browser
Application execution is secure
| Secure execution through:
Elimination of unsafe instructions
Safe branch jumps
Safe function calls and function returns
16
| To provide secure environment, all instructions are bundled.
| Before jump, always check. Even for function calls!
| Solution: Simple! detect a pattern and store the return address in a stack
| Lessons: Security solutions can create new branch prediction problems.
Call PC
return
NaCl RET Detector NaCl-RAS
Indirect Predictor
nacl_ret ? Pop
Prediction
reg_ret ? Pop
Pop/Mask reg_call ? Push
18
| Still Lots of challenges/Issues in Mobile Platforms
Energy, efficiency, security etc.
| But that’s not the end! Energy, Security, Reliability, and performance
& Thank Yale!
19
| Students and Collaborators: Dilan Manatunga, Hyojong Kim, Nagesh B Lakshminarayana, Pranith Kumar, Hyongyeol Lim, Gi-ho Park, Saibal Mukhopadhyay
