Customizable Embedded System Architectures

1
Customizable Embedded System Architectures

• Peter Petrov and Alex Orailoglu

University of California, San Diego
2
Embedded Processors Market

• Embedded processors occupy more than 90 of the
  entire processor market
• A large number of electronic products require
  high-end 32/64-bits embedded processors

3
Embedded Processors Market

• Embedded processors occupy more than 90 of the
  entire processor market
• A large number of electronic products require
  high-end 32/64-bits embedded processors

4
Application Requirements

• Design cost

General embedded processor architectures
introduced to satisfy these constraints!

• Time-to-market

• Flexibility

5
Application Requirements

• Design cost

General embedded processor architectures
introduced to satisfy these constraints!

• Time-to-market

• Flexibility

Processor architecture

• Deterministic Performance

• Power consumption

• Performance

6
New Architectural Paradigm
Design costTime-to-marketFlexibilityDeterminism
PowerPerformance
Design costTime-to-marketFlexibilityDeterminism
PowerPerformance
Design costTime-to-marketFlexibilityDeterminism
PowerPerformance
7
New Architectural Paradigm
DeterminismPowerPerformance Design
costTime-to-marketFlexibility
FPGA
ASC?P
?P
ASIC

• Application-Specific Customizable Embedded
  Processor
• Helps preserve the benefits of generality
• Alleviates the drawbacks of general-purpose
  processors

8
Static vs. Dynamic Optimizations in General
Purpose Processors
ApplicationInformation
Application
Micro-Architecture
Hardware for dynamic resolution
CompilerOptimizations
ArchitecturalOptimizations
9
Dynamically Customizable Embedded Processors
Application
Micro-Architecture
ISA
Hardware for dynamic resolution
ASCL
Execution Resources
ASCL Application Specific Customization Logic

• Compilers for static extraction
• Architectural Runtime incorporation
• ASCL shapes the processor by matching compiler
  information to microarchitecture

• Provides deterministic information about
  application regularities
• Restricts the domain of possible application
  behaviors

10
Microarchitectural Customizations

• Use of application knowledge in
  microarchitectural modules
• Power
• Performance
• Determinism
• Reprogrammable customization hardware
• Post-manufacturingre-customizations
• Large manufacturing volumes

FU2
FU1
Program RAM
Application
?P
11
Information Transfer and Hardware Support
Loop A
ASCL
App Information
Loop B
Loop C
Application-Specific ?P
Application
Special registers or tables

• Application hot spots targeted
• Application information loaded into special
  hardware tables/registers, providing
  reprogrammable implementation
• Information transfer either by software or system
  setup

12
Unified Customizable Architecture
Data Memory
ALU
A unified, dynamically customizable embedded
processor architecture
13
Application Knowledge BenefitsPrecise
application knowledge
Application
Micro-Architecture
ISA
CachesBranch Pred., etc.
FU2

FU1

• Statistically based methods normally used to
  infer application properties
• Power expensive microarchitectural components
• Highly sub-optimal performance
• Unpredictable execution time

14
Application Knowledge BenefitsPrecise
application knowledge
Application
Micro-Architecture
ISA
CachesBranch Pred., etc.
ASCL
FU2

FU1

• Statistically based methods normally used to
  infer application properties
• Precise application knowledge used instead
  through the ASCL
• Application regularities readily available for
  utilization gt Scaled down and power efficient
  uArchitectural components
• Deterministic execution time achieved

15
Application Knowledge BenefitsRefined program
behavior
ISA
Micro-Architecture
Worst case assumption for the program execution
P4
FU2

P5
FU1
A large set of potential programs!

• Worst case execution scenario assumed in general
  purpose processor

16
Application Knowledge BenefitsRefined program
behavior
ISA
Micro-Architecture
Considering a single program segment only!
ASCL
FU2

FU1
A large set of potential programs!

• Worst case execution scenario assumed in general
  purpose processors
• Application knowledge refines the domain of all
  possible states
• Redundant hardware activities removed gt Power
  savings

17
Conclusions

• A customizable processor architecture defined
• In-field recustomization
• High volumes due to fixed-silicon architecture
• A unified architecture for diverse sets of tasks
• The adaptive architecture provides flexibility,
  high utilization, and low power for an ever
  increasing and diverse set of applications
• Experimentally verified orders of magnitude
  improvements