ChipNeMo
NVIDIA's domain-adapted chip-design LLMs: LLaMA 2 continued-pretrained on 20B+ tokens of chip data, matching GPT-4 at 70B.
About
ChipNeMo is NVIDIA's recipe for domain-adapting LLMs to chip design, presented at ICCAD 2023. Rather than relying on a giant general model, the team took LLaMA 2 and applied a full adaptation stack: a domain-adaptive tokenizer, continued pretraining on over 20 billion tokens of chip-design code and documents, instruction alignment with domain instructions, and a domain-adapted retrieval model. They built 7B, 13B, and 70B variants and found the adaptation delivered up to 5x model-size reduction at matched quality.
The payoff was production-relevant: ChipNeMo-70B reportedly beat GPT-4 on two of three internal use cases, an engineering-assistant chatbot and EDA script generation, with competitive results on bug summarization and analysis. The weights were never publicly released, so this entry is a landmark reference for domain adaptation methodology rather than a downloadable model. NVIDIA published tutorial playbooks for replicating the DAPT pipeline in the NeMo framework.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Internal NVIDIA evals: ChipNeMo-70B surpasses GPT-4 on engineering chatbot and EDA script generation (provider-reported, ICCAD 2023 paper)
- Up to 5x model-size reduction at similar or better quality vs base LLaMA 2 (provider-reported)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (metrics tier, metrics-v1). Evidence: peer-reviewed ICCAD 2023 paper (arXiv:2311.00176) from NVIDIA Research with detailed methodology and three production use cases; claims independently summarized by third-party technical surveys. Strengths: rigorous, quantified domain-adaptation results from a credible industrial lab. Docked for: (1) all evaluations are internal to NVIDIA with no public replication; (2) no weights, no API, and no commercial path — it is a methodology reference, not a usable product.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ICCAD 2023 peer-reviewed; 5x size-reduction claim summarized by third-party surveys
- · All evals internal to NVIDIA; methodology reference, not a usable model