ChipNeMo: Domain-Adapted LLMs for Chip Design
read the original abstract
ChipNeMo aims to explore the applications of large language models (LLMs) for industrial chip design. Instead of directly deploying off-the-shelf commercial or open-source LLMs, we instead adopt the following domain adaptation techniques: domain-adaptive tokenization, domain-adaptive continued pretraining, model alignment with domain-specific instructions, and domain-adapted retrieval models. We evaluate these methods on three selected LLM applications for chip design: an engineering assistant chatbot, EDA script generation, and bug summarization and analysis. Our evaluations demonstrate that domain-adaptive pretraining of language models, can lead to superior performance in domain related downstream tasks compared to their base LLaMA2 counterparts, without degradations in generic capabilities. In particular, our largest model, ChipNeMo-70B, outperforms the highly capable GPT-4 on two of our use cases, namely engineering assistant chatbot and EDA scripts generation, while exhibiting competitive performance on bug summarization and analysis. These results underscore the potential of domain-specific customization for enhancing the effectiveness of large language models in specialized applications.
This paper has not been read by Pith yet.
Forward citations
Cited by 25 Pith papers
-
AgentDSE: Reasoning-Augmented Architectural Design Space Exploration
AgentDSE uses an LLM agent in a simulator-in-the-loop setup to achieve competitive or superior architectural designs with up to 100x fewer evaluations than traditional black-box optimization methods.
-
PDAGENT-BENCH: Characterizing, Grounding, and Architecting LLM Agents for VLSI Physical Design
PDAGENT-BENCH is a new benchmark suite with 353 curated problems and an agentic workflow framework for evaluating LLM/VLM agents across five capability dimensions in VLSI physical design.
-
Alpha-RTL: Test-Time Training for RTL Hardware Optimization
TTT-RTL performs per-design test-time RL on an LLM policy with EDA-derived PPA rewards and an adaptive KL controller, reducing geometric-mean PPA product by 65.1% on RTLLM v2.0 and ADP by 59.4% on an industrial FPU unit.
-
Bridging the Last Mile of Circuit Design: PostEDA-Bench, a Hierarchical Benchmark for PPA Convergence and DRC Fixing
PostEDA-Bench shows LLM agents achieve reasonable performance on synthetic DRC-Essential and single-objective PPA tasks but only 36.66% success on DRC-Reasoning and 20% on multi-objective PPA tasks.
-
Bridging the Last Mile of Circuit Design: PostEDA-Bench, a Hierarchical Benchmark for PPA Convergence and DRC Fixing
PostEDA-Bench shows LLM agents succeed reasonably on basic DRC and single-objective PPA tasks but struggle on practical DRC reasoning (best 36.66% success) and multi-objective PPA (best 20% success).
-
DiagramNet: An End-to-End Recognition Framework and Dataset for Non-Standard System-Level Diagrams
DiagramNet supplies a new multimodal dataset and progressive training pipeline with decoupled multi-agent workflow, allowing a 3B model to outperform GPT-5, Claude-Sonnet-4, and Gemini-2.5-Pro by over 2x on system-lev...
-
HAVEN: Hybrid Automated Verification ENgine for UVM Testbench Synthesis with LLMs
HAVEN combines LLM agents for planning and gap analysis with protocol-specific templates and a custom DSL to generate correct UVM testbenches, achieving 100% compilation success, 90.6% code coverage, and 87.9% functio...
-
TimingLLM: A Two-Stage Retrieval-Augmented Framework for Pre-Synthesis Timing Prediction from Verilog
TimingLLM uses a fine-tuned LLM to generate structural timing cues from Verilog followed by a retrieval-augmented regressor with a learned steering vector to predict WNS and TNS with R values of 0.91 and 0.97.
-
LEGO: An LLM Skill-Based Front-End Design Generation Platform
LEGO extracts 42 circuit skills from open-source projects to enable composable LLM-based front-end design, raising Pass@1 to 0.805 on challenging Verilog problems.
-
Structural Verification for Reliable EDA Code Generation without Tool-in-the-Loop Debugging
Structural dependency graphs and staged pre-execution verification raise LLM-based EDA code pass rates to 82.5% (single-step) and 70-84% (multi-step) while halving tool calls by catching dependency violations before runtime.
-
ATLAAS: Automatic Tensor-Level Abstraction of Accelerator Semantics
ATLAAS automatically converts RTL-extracted bit-level accelerator semantics into tensor-level ISA specs via an 8-pass MLIR pipeline, enabling automated compiler backend generation for designs like Gemmini and VTA.
-
Arch: An AI-Native Hardware Description Language for Register-Transfer Clocked Hardware Design
Arch is a new AI-native HDL that uses parameterized clock/reset types and built-in hardware primitives to enable type-safe, AI-generatable register-transfer designs compiling to SystemVerilog with automatic formal properties.
-
FVRuleLearner: Operator-Level Reasoning Tree (OP-Tree)-Based Rules Learning for Formal Verification
FVRuleLearner introduces an Operator Reasoning Tree to learn operator-specific rules that improve natural-language to SystemVerilog assertion generation, raising syntax correctness by 3.95% and functional correctness ...
-
Verilog-Evolve: Feedback-Driven and Skill-Evolving Verilog Generation
Verilog-Evolve uses executable feedback from simulation, synthesis, timing, and GEMM metrics to refine LLM-generated Verilog and evolves skills across tasks, improving functional success and downstream hardware qualit...
-
RefEvo: Agentic Design with Co-Evolutionary Verification for Agile Reference Model Generation
RefEvo achieves 95% pass rate on 20 hardware modules for SystemC reference model generation using dynamic multi-agent planning, co-evolutionary verification, and spec anchoring, with 71% token reduction.
-
COEVO: Co-Evolutionary Framework for Joint Functional Correctness and PPA Optimization in LLM-Based RTL Generation
COEVO unifies correctness and multi-objective PPA optimization in a single evolutionary loop for LLM RTL generation, reporting 97.5% and 94.5% Pass@1 on VerilogEval/RTLLM benchmarks plus best PPA on 43 of 49 designs.
-
Graph Computation Meets Circuit Algebra: A Task-Aligned Analysis of Graph Neural Networks for Electronic Design Automation
GNNs succeed in EDA when their propagation, aggregation, and supervision match the native algebra of each circuit task, such as max-plus recurrences for timing or hypergraph penalties for placement.
-
ChipLingo: A Systematic Training Framework for Large Language Models in EDA
ChipLingo trains LLMs on EDA data via corpus construction, domain-adaptive pretraining, and RAG scenario alignment, reaching 59.7% accuracy with an 8B model and 70.02% with a 32B model on a new internal EDA benchmark.
-
LEGO: An LLM Skill-Based Front-End Design Generation Platform
LEGO extracts 42 standardized circuit skills from 11 open-source projects into a plug-and-play platform that raises Pass@1 from 0 to 0.805 on 41 hard VerilogEval v2 problems.
-
ORFS-agent: Tool-Using Agents for Chip Design Optimization
ORFS-agent uses LLM agents to tune parameters in chip design flows, improving geometric-mean wirelength, clock period, and co-optimization objectives by up to 2.7% over OR-AutoTuner with 40% fewer iterations on ASAP7 ...
-
StepPRM-RTL: Stepwise Process-Reward Guided LLM Fine-Tuning for Enhanced RTL Synthesis
StepPRM-RTL combines stepwise process-reward modeling, RAFT, and MCTS to fine-tune LLMs for RTL code, claiming over 10% gains in functional correctness on Verilog/VHDL benchmarks.
-
From AGI to ASI
The paper characterizes ASI and examines scaling, paradigm shifts, recursive self-improvement, and multi-agent collectives as routes from AGI to ASI, together with frictions and open questions about acceleration.
-
Physics-informed generative AI for semiconductor manufacturing: Enforcing hard physical constraints in generative models by construction
Perspective arguing that physics-informed generative models enforcing hard constraints by construction outperform post-hoc filtering for semiconductor manufacturing and similar physical systems.
-
Evaluating Hallucinations in Domain-Adapted Large Language Models
Fine-tuning Llama-2 on a small domain-specific dataset yields high memorization but near-zero reasoning on newly introduced entities, suggesting fine-tuning alone is insufficient for knowledge injection.
-
Focus Session: Hardware and Software Techniques for Accelerating Multimodal Foundation Models
The paper compiles hardware-software co-design techniques including mixed-precision quantization, structural pruning, speculative decoding, and transformer accelerators to speed up multimodal foundation models, with e...
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.