Title: An AlgorithmArchitecture CoDesign Methodology for Wireless Digital Receivers
1An Algorithm/Architecture Co-Design Methodology
for Wireless Digital Receivers
- Ning Zhang
- Bob Brodersen
- BWRC Winter Retreat, January 2000
2Wireless Communication System Trend
- Increasing demands
- High capacity
- High data rate
- with constraints
- Portability
- low energy consumption
- small form factor
- Fast time-to-market
3The Gap Between Reality and Theory
Commercial Systems (Proxim WLAN)
New Theoretical Results (Lucent V-BLAST)
Spectral Efficiency
Single-user 1-2 bps/Hz Multi-user 0.3 bps/Hz
12-40 bps/Hz
Implementation
1/10 of real-time, 650 Kbps 4 TMS32C40 DSP 7 W
1.6 Mbps 750 mW(includes analog)
Lower Power High Performance at high data rate
4Problem Statement
- How to facilitate low power digital
implementation of high performance wireless
communication algorithms?
Algorithm
Establish an algorithm/architecture co-design
methodology that enables efficient and effective
algorithm and architecture explorations.
Architecture
Digital Signal Processor
Dedicated Hardware
Reconfigurable Hardware
5Methodology Outline
Algorithm
Algorithm Characteristic Evaluation
Architecture
Digital Signal Processor
Dedicated Hardware
Reconfigurable Hardware
6Algorithm Evaluation
Performance and Implementation Complexity
- Theoretical capacity
- Algorithm performance
- e.g., convergence speed of adaptive algorithm
- Operation count
- Operation breakdown
- e.g., MULT, DIV, SQRT, ALU and MEM
- Sensitivity to analog circuit nonidealities
- Numerical robustness with finite precision
- Structure
- e.g., locality, regularity and concurrency
Performance
Implementation Complexity
7Methodology Outline
Algorithm
Architectural Implementation Estimation
Architecture
Digital Signal Processor
Dedicated Hardware
Reconfigurable Hardware
8Architectural Implementation Estimation
- Digital Signal Processor
- Kernel benchmark approach
- Start from existing architectures
- compiled assembly codes instruction level power
characterizations - Evaluate relative improvements for extensions of
architectures - Retargetable estimation Ghazal
- Reconfigurable Hardware
- Pleiades Wan
- Stream-based dataflow architecture Kienhuis
- Dedicated Hardware
9Low Power Design Techniques
- Identify and Estimate Potential Improvements
- Accumulate Design Knowledge
- Algorithm
- Reduce operation complexity
- Optimize computation structure
- Architecture
- Optimize fundamental components
- Reduce overhead components
10Methodology Outline
Algorithm
Fast and Predictable Implementation Path
Architecture
Digital Signal Processor
Dedicated Hardware
Reconfigurable Hardware
11Front-End Design for Dedicated Hardware
Fast and Predictable Implementation
- Use block diagram based algorithm specification
-
- Preserve algorithm structure to establish
connections between algorithm, architecture and
physical levels
increase predictability and shorten feedback loop
- Design at high level by reusing building blocks
increase productivity
12Module Based Approach
Algorithm
Recommendation for Modifications and
Transformations
Pre-Characterized Module Library with PDA Models
Architecture
Data Flow (construct with parameterized modules)
Control Flow
Estimation and Evaluation
(synthesize to standard cells)
(generate C code)
Netlist and Floor Plan of Macro Modules
Standard Cells
Implementation on Processor
Physical Design Information
Back-end Design flow
13Design Example
A
.
m
- solve to minimize
- decompose A to an unitary and an upper
triangular matrix - get A-1 (mn)
n
- Applications
- interference suppression and diversity combining
for multiple antenna systems - multi-user detection for CDMA systems
- recursive least squares filters
- any algorithm requiring matrix inversion
14Hardware Design Methodology
- Design Space Exploration
- System level
- performance and implementation complexity
- Algorithm level
- numerical and structural properties
- Architecture level
- parallel and pipeline schemes (speed/area/energy)
- Arithmetic level
- MAC/DIV/SQRT or CORDIC
- Design Framework
- High level modeling and estimation
- Optimizations across all levels
- System/Hardware codesign