Skip to content

Design Space Exploration

Overview

Design space exploration (DSE) is the systematic process of evaluating alternative design choices to find solutions that best meet a set of objectives and constraints. In HW/SW co-design, the design space encompasses all possible combinations of partitioning, mapping, scheduling, and implementation decisions.

DSE is essential because embedded system design involves navigating trade-offs between competing objectives — no single design is typically best on all metrics.

The Design Space

The design space for an embedded system includes choices at multiple levels:

Level Design Choices
Architecture Platform selection, number and type of processing elements
Partitioning Which tasks are implemented in HW vs SW
Mapping Which tasks are assigned to which resources
Scheduling In what order tasks execute, priority assignment
Implementation Algorithm variants, data representations, communication protocols

The total number of candidate designs grows combinatorially with the number of choices at each level, making exhaustive evaluation infeasible for all but the smallest systems.

Objectives and Constraints

DSE aims to optimise one or more objectives while satisfying constraints:

Common objectives:

  • Minimise execution time / latency
  • Minimise energy or power consumption
  • Minimise hardware area or cost
  • Maximise throughput

Common constraints:

  • Real-time deadlines must be met
  • Memory capacity must not be exceeded
  • Power budget must not be exceeded
  • Hardware area must fit the target device

Note

Objectives often conflict — for example, reducing latency may require more hardware resources (higher cost), while minimising cost may increase execution time.

Multi-Objective Optimisation

When multiple objectives are in play, there is generally no single "best" design. Instead, DSE identifies the set of Pareto-optimal solutions.

A design \(A\) dominates design \(B\) if \(A\) is at least as good as \(B\) on all objectives and strictly better on at least one. The Pareto front is the set of all non-dominated designs.

Cost
 ^
 |  x
 |    x         x = candidate designs
 |      x       o = Pareto-optimal designs
 |   o    x
 |     o
 |       o
 |         o
 +-----------------> Latency

The designer then selects from the Pareto front based on application priorities and constraints.

DSE Methods

Exhaustive Enumeration

Evaluate every possible design. Guarantees finding the optimal solution but only feasible for very small design spaces.

  • Complexity: Product of choices at each level (often exponential)
  • Use case: Small problems, validation of heuristic results

Use guided search strategies to explore the design space efficiently:

  • Simulated annealing: Random perturbations with a cooling schedule that gradually reduces the acceptance of worse solutions
  • Genetic algorithms: Population-based search using selection, crossover, and mutation operators
  • Tabu search: Local search with memory to avoid revisiting recently explored designs

Analytical Models

Use mathematical models to estimate performance without full simulation:

  • Worst-case execution time (WCET) analysis for timing
  • Energy models based on activity factors and switching capacitance
  • Queuing models for communication and resource contention

Simulation-Based Evaluation

Run detailed simulations of candidate designs:

  • Cycle-accurate simulation for timing accuracy
  • Power simulators for energy estimation
  • Co-simulation of HW and SW components together

Tip

In practice, DSE often combines fast analytical models for initial pruning with detailed simulation for evaluating the most promising candidates.

DSE Flow

A typical design space exploration flow:

  1. Define the design space — identify all design choices and their ranges
  2. Define objectives and constraints — what to optimise, what must be satisfied
  3. Sample candidate designs — exhaustive, random, or guided by a search algorithm
  4. Evaluate each candidate — using models, simulation, or estimation
  5. Identify Pareto-optimal solutions — filter dominated designs
  6. Select the final design — based on application priorities

Summary

Concept Description
Design space All possible combinations of design choices
Pareto optimality No design is better on all objectives simultaneously
Pareto front The set of non-dominated designs
Exhaustive search Optimal but exponential complexity
Heuristic search Efficient but no optimality guarantee
Analytical models Fast estimation for pruning the space
Simulation Detailed evaluation of promising candidates