Agent Factories for High Level Synthesis: How Far Can General-Purpose Coding Agents Go in Hardware Optimization?

5 Steps

1. 1

   Define Target Algorithm: Choose a high-level algorithmic specification (e.g., C/C++ function) that is suitable for hardware implementation via High-Level Synthesis (HLS). This will serve as the initial input for your AI agents.

2. 2

   Configure General-Purpose LLM Agents: Set up access to a powerful Large Language Model (LLM) and instantiate client connections. Define distinct roles and initial prompts for each agent (e.g., 'HLS Designer', 'Optimizer', 'Evaluator') to guide them without explicit hardware-specific training.

3. 3

   Design the Agent Factory Orchestration: Implement a 'two-stage pipeline' or a multi-agent coordination mechanism. This 'agent factory' will manage the construction, communication, and task assignment among your agents, guiding them through iterative optimization cycles.

4. 4

   Execute Optimization Cycles: Initiate the agents to generate and refine hardware descriptions (e.g., Verilog, VHDL, or optimized C/C++ for HLS). Allow agents to iterate, propose modifications, and apply optimization strategies based on their general problem-solving capabilities.

5. 5

   Evaluate Hardware Metrics: Use HLS tools (e.g., Vitis HLS, Intel HLS) to synthesize and analyze the generated hardware designs. Evaluate key metrics such as performance (latency, throughput), resource utilization (area), and power consumption to assess the agents' optimization effectiveness.