How nvlabs/ace-rtl works
ACE-RTL is an agentic RTL generation, verification, and iterative-repair system. The principal workflow starts with a user-selected task, runs generation and verification, uses tool feedback to guide reflection and coordination, and returns artifacts and results. Its runtime also includes a native CVDP evaluator and shared LLM backend. The graph preserves the README’s Generator/Reflector/Coordinator workflow and major benchmark integrations; edges are limited to README statements and SOURCE INDEX references, so unsampled implementation details are not inferred.
Main entry points: cli.py Native runner · agent.py Agent runtime · llm_centric_debugger.py LLM debugger · comprehensive_error_analyzer.py Error analysis
Components 26
Agent orchestration
- Agent CLI
skills/ace-rtl/scripts/ace_rtl_agent/main.py - Agent runtime
skills/ace-rtl/scripts/ace_rtl_agent/agent.py - Iteration loop
skills/ace-rtl/scripts/ace_rtl_agent/agent_loop.py - Run reporting
skills/ace-rtl/scripts/ace_rtl_agent/agent_reporting.py - Role coordinator
skills/ace-rtl/scripts/ace_rtl_agent/fresh_start_coordinator.py
RTL reasoning
- RTL generator
skills/ace-rtl/scripts/ace_rtl_agent/rtl_generator.py - Focused debugger
skills/ace-rtl/scripts/ace_rtl_agent/focused_debugger.py - LLM debugger
skills/ace-rtl/scripts/ace_rtl_agent/llm_centric_debugger.py - Error analysis
skills/ace-rtl/scripts/ace_rtl_agent/comprehensive_error_analyzer.py - Failure extraction
skills/ace-rtl/scripts/ace_rtl_agent/primary_failure_extractor.py - RTL rule advice
skills/ace-rtl/scripts/ace_rtl_agent/rtl_rule_advisor.py - Testbench analysis
skills/ace-rtl/scripts/ace_rtl_agent/testbench_analyzer.py
LLM services
- LLM client
skills/ace-rtl/scripts/ace_rtl_agent/utils/llm_client.py - ACE model adapter
skills/ace-rtl/scripts/ace_rtl_agent/utils/ace_model_client.py - LLM backend
skills/ace-rtl/scripts/ace_rtl_agent/llm_backend.py
Benchmark evaluation
- CVDP integration
skills/ace-rtl/scripts/ace_rtl_agent/utils/cvdp_environment.py - VerilogEval integration
skills/ace-rtl/scripts/ace_rtl_agent/utils/verilogeval_integration.py - ChipBench integration
skills/ace-rtl/scripts/ace_rtl_agent/utils/chipbench_integration.py - Native runner
skills/ace-rtl/scripts/ace_cvdp_native/cli.py - Task configuration
skills/ace-rtl/scripts/ace_cvdp_native/config.py - Run processes
skills/ace-rtl/scripts/ace_cvdp_native/processes.py - CVDP harness
skills/ace-rtl/scripts/ace_cvdp_native/harness.py - Result diagnostics
skills/ace-rtl/scripts/ace_cvdp_native/diagnostics.py - SystemVerilog helpers
skills/ace-rtl/scripts/ace_cvdp_native/sv.py
Other components
- Task user
- NVIDIA API
Connections 29
- Task user → Agent CLI: submits task
- Agent CLI → Agent runtime: starts runtime
skills/ace-rtl/scripts/ace_rtl_agent/main.py - Agent runtime → Iteration loop: delegates iteration
skills/ace-rtl/scripts/ace_rtl_agent/agent.py - Agent runtime → Run reporting: delegates reporting
skills/ace-rtl/scripts/ace_rtl_agent/agent.py - Agent runtime → RTL generator: uses generation
skills/ace-rtl/scripts/ace_rtl_agent/agent.py - Agent runtime → CVDP integration: uses evaluation
skills/ace-rtl/scripts/ace_rtl_agent/agent.py - Agent runtime → VerilogEval integration: uses evaluation
skills/ace-rtl/scripts/ace_rtl_agent/agent.py - Agent runtime → ChipBench integration: uses integration
skills/ace-rtl/scripts/ace_rtl_agent/agent.py - Iteration loop → Role coordinator: coordinates roles
- Iteration loop → Focused debugger: reflects on failures
- LLM debugger → Focused debugger: uses debugging
skills/ace-rtl/scripts/ace_rtl_agent/llm_centric_debugger.py - LLM debugger → Testbench analysis: analyzes tests
skills/ace-rtl/scripts/ace_rtl_agent/llm_centric_debugger.py - LLM debugger → LLM client: requests analysis
skills/ace-rtl/scripts/ace_rtl_agent/llm_centric_debugger.py - LLM debugger → ACE model adapter: requests model call
skills/ace-rtl/scripts/ace_rtl_agent/llm_centric_debugger.py - Error analysis → Failure extraction: extracts failure
skills/ace-rtl/scripts/ace_rtl_agent/comprehensive_error_analyzer.py - Error analysis → LLM client: requests analysis
skills/ace-rtl/scripts/ace_rtl_agent/comprehensive_error_analyzer.py - Error analysis → RTL rule advice: consults rules
skills/ace-rtl/scripts/ace_rtl_agent/comprehensive_error_analyzer.py - LLM client → LLM backend: calls backend
skills/ace-rtl/scripts/ace_rtl_agent/utils/llm_client.py - ACE model adapter → LLM backend: calls backend
skills/ace-rtl/scripts/ace_rtl_agent/utils/ace_model_client.py - LLM backend → NVIDIA API: sends requests
- Native runner → Task configuration: loads configuration
skills/ace-rtl/scripts/ace_cvdp_native/cli.py - Native runner → Run processes: uses process helpers
skills/ace-rtl/scripts/ace_cvdp_native/cli.py - Native runner → CVDP harness: runs evaluation
skills/ace-rtl/scripts/ace_cvdp_native/cli.py - Native runner → SystemVerilog helpers: uses RTL helpers
skills/ace-rtl/scripts/ace_cvdp_native/cli.py - Native runner → Result diagnostics: summarizes results
skills/ace-rtl/scripts/ace_cvdp_native/cli.py - CVDP harness → Task configuration: reads settings
skills/ace-rtl/scripts/ace_cvdp_native/harness.py - CVDP harness → Run processes: uses process helpers
skills/ace-rtl/scripts/ace_cvdp_native/harness.py - Native runner → Role coordinator: invokes agent roles
- Native runner → CVDP integration: evaluates benchmark