Skip to content

Electronic Design Automation

Graph-Based RTL Generation

Challenge: Translating abstract graph-based models (like data flow graphs or process communication graphs) directly into structurally correct RTL code.

Year Venue Authors Title Tags P E N
2013 DAC Columbia A Method to Abstract RTL IP Blocks into C++ Code and Enable High-Level Synthesis process communication graph; I/O port loop unrolling; HLS design space expansion
2024 ISEDA UESTC GraphRTL: an Agile Design Framework of RTL Code from Data Flow Graphs graph error detection kernel; DFS based graph equivalent reconstruction; template/Scala based DFG and CFG merging

Dynamically Scheduled HLS and Dataflow Circuits

Challenge: HLS-generated dataflow circuits force successive instances of the same operation to execute strictly in program order due to the absence of token tags, preventing exploitation of parallelism analogous to out-of-order superscalar processors—particularly harmful when long-latency operations or high-II inner loops create pipeline bubbles.

Solution: Introduce selective token tagging and alignment components (TAGGER, ALIGNER, UNTAGGER) into untagged dataflow circuits at formally derived minimal positions, enabling out-of-order execution of targeted operations while preserving correctness for loop-carried dependencies and memory hazards.

Year Venue Authors Title Tags P E N
2024 FPGA EPFL Survival of the Fastest: Enabling More Out-of-Order Execution in Dataflow Circuits TAGGER/ALIGNER/UNTAGGER components for selective token tagging in untagged dataflow circuits; dirty-node and unaligned-edge sets for minimal ALIGNER positioning; noncyclic/cyclic control-flow clustering to prevent alignment count mismatch 5 4 4

Spatial Accelerator Generation

Challenge: Building parameterized compilation frameworks to rapidly synthesize, evaluate, and deploy diverse spatial architectures (e.g., systolic arrays) while decoupling algorithm mapping from hardware generation.

Year Venue Authors Title Tags P E N
2020 ISCA UCLA DSAGEN: Synthesizing Programmable Spatial Accelerators architecture description graph (ADG); solution-repairing spatial scheduler to avoid redundant re-compilation; composable decoupled spatial primitives for approximating diverse accelerators 4 3 4
2021 ASPLOS Cornell A compiler infrastructure for accelerator generators a split representation combining a high-level control flow language with a hardware-like structural language; pass-based compiler; systolic array generator; live-range-based register-sharing 4 3 3
2021 DAC UCB Gemmini: Enabling Systematic Deep-Learning Architecture Evaluation via Full-Stack Integration exploring diverse systolic array dimensions and dataflows; RoCC(Rocket Custom Coprocessor) integration with unified virtual memory support; FPGA-accelerated full-system simulation 4 5 4

LLM-Driven RTL Generation

Challenge: Overcoming the lack of domain knowledge and strict syntax adherence in generic Large Language Models (LLMs) to generate functionally correct and synthesizable Verilog code.

Agentic RTL Code Generation

Challenge: Single-shot LLM prompting loses circuit details and lacks domain-specific feedback loops for correcting syntax and functional RTL errors.

Solution: Equip RTL coding agents with task decomposition, simulator feedback, graph-based circuit context retrieval, and waveform-aware debugging tools.

Year Venue Authors Title Tags P E N
2024 arXiv UCSD MAGE: A Multi-Agent Engine for Automated RTL Code Generation multi-agent; high-temperature sampling and ranking; verilog-state checkpoint debugging
2025 AAAI NVIDIA VerilogCoder: Autonomous Verilog Coding Agents with Graph-based Planning and Abstract Syntax Tree (AST)-based Waveform Tracing Tool Task and Circuit Relation Graph; AST-based waveform tracing tool; task-driven circuit relation graph retrieval agent; code agent and debug agent split; VerilogEval-Human v2 single-run flow 4 3 4

RTL Model Training and Repair Data

Challenge: Verilog code models are bottlenecked by low-quality synthetic training data, non-textual RTL specifications, and small but recurring functional mistakes.

Solution: Curate Verilog-specific training data with code augmentation, self-reflection, correct-by-construction non-textual examples, and targeted code repair cases.

Year Venue Authors Title Tags P E N
2024 ICCAD PKU OriGen: Enhancing RTL Code Generation with Code-to-Code Augmentation and Self-Reflection self-reflection mechanism; dataset augmentation methodology; VerilogFixEval benchmark 2 4 2
2025 ICLR NVIDIA CraftRTL: High-quality Synthetic Data Generation for Verilog Code Models with Correct-by-Construction Non-Textual Representations and Targeted Code Repair correct-by-construction KMap/FSM/waveform data; VerilogEval-NonText subset; checkpoint error report construction; targeted code repair injection; SDG-CC-Repair training mixture 4 3 3

Higher-Level RTL Code Generation

Challenge: raise the level of abstraction to specifically target the construction of applications

Year Venue Authors Title Tags P E N
2022 PLDI Cornell University PDL: A High-Level Hardware Design Language for Pipelined Processors one-instruction-at-a-time semantics; hardware description language; hazard locks; speculation API; guaranteed-correct pipelining 3 3 3
2025 ISCA KAUST Assassyn: A Unified Abstraction for Architectural Simulation and Implementation unified simulation and RTL generation; asynchronous event-driven pipeline abstraction; function-as-stage programming model; transposed trace-waveform alignment 3 3 3

RTL Code Generation Benchmarks

Challenge: The lack of standardized benchmarks for evaluating RTL code generation tools.

Year Venue Authors Title Tags P E N
2023 DATE NYU Benchmarking Large Language Models for Automated Verilog RTL Code Generation verilog code training corpus; multi-level verilog coding problems for analysis
2025 arXiv NVIDIA Comprehensive Verilog Design Problems: A Next-Generation Benchmark Dataset for Evaluating Large Language Models and Agents on RTL Design and Verification CVDP benchmark; 783 human-authored Verilog problems; Agentic and Non-Agentic JSONL formats; Dockerized EDA tool harness; category-level failure clustering 4 3 3

LLM-Driven Hardware Design Agents

Challenge: General-purpose LLM agents do not directly handle hardware design workflows that require domain knowledge, EDA tools, timing reports, layout checks, and task-specific collaboration patterns.

Solution: Use configurable graph-based task solving with single-AI, multi-AI, and multi-modality agents connected to hardware-specific tools, memory, and knowledge databases.

Year Venue Authors Title Tags P E N
2025 VLSI NVIDIA Marco: Configurable Graph-Based Task Solving and Multi-AI Agents Framework for Hardware Design dynamic and static configurable task graph; sub-task node agent configuration; MCMM timing analysis agent; DRC Coder multi-modality agent; timing path debug static task graph 3 3 2

Physical Design

Challenge: Optimizing the physical layout (placement and routing) of circuits on chips to meet timing, power, and area constraints, especially for large-scale heterogeneous architectures.

Year Venue Authors Title Tags P E N
2019 ICCAD UT Austin elfPlace: Electrostatics-based Placement for Large-Scale Heterogeneous FPGAs electrostatics-based placement for heterogeneous FPGAs; augmented Lagrangian formulation; preconditioning technique; normalized subgradient multiplier updating; unified instance area adjustment 4 4 4
2023 ICCAD TAMU Systolic Array Placement on FPGAs Region-wise Sweep in Alternating Direction (R-SAD); partition enumeration and pruning; regularity-driven placement; DSP column placement optimization 3 4 3

HLS-aware Physical Design

Challenge: Long compilation time for HLS designs due to the inherently sequential and complex physical implementation (place & route) flow.

Solution: Leverage HLS dataflow properties to enable a parallel physical implementation flow.

Year Venue Authors Title Tags P E N
2022 FPGA UCLA RapidStream: Parallel Physical Implementation of FPGA HLS Designs split compilation; HLS-physical co-design; anchor registers; timing isolation; parallel placement and routing; stitching 4 4 4

Logic Synthesis & Technology Mapping

Challenge: Mapping behavioral or structural RTL descriptions to complex, architecture-specific FPGA or ASIC primitives (like DSPs) efficiently and correctly, avoiding heuristic fallbacks.

Solution: Employ automated reasoning, such as sketch-guided program synthesis, to systematically explore the logic configuration space.

FPGA Technology Mapping

Year Venue Authors Title Tags P E N
2024 ASPLOS UW FPGA Technology Mapping Using Sketch-Guided Program Synthesis Lakeroad program synthesis-based technology mapper; architecture-independent sketch templates; automated semantics extraction from vendor HDL models; Rosette-based SMT solver translation 3 4 3