Reproduce “TPU-Gen: LLM-Driven Custom Tensor Processing Unit Generator” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Deepak Vungarala, Mohammed E. Elbtity, Sumiya Syed, Sakila Alam, Kartik Pandit, Arnob Ghosh, Ramtin Zand, Shaahin Angizi
Researchers from the New Jersey Institute of Technology and the University of South Carolina developed TPU-Gen, the first framework leveraging Large Language Models (LLMs) to automate the generation of custom Tensor Processing Units (TPUs), including approximate computing variants. Integrating Retrieval-Augmented Generation (RAG) and a curated dataset, TPU-Gen consistently produced designs that met Power, Performance, and Area (PPA) targets, outperforming manual optimization efforts with an average reduction of 92% in area and 96% in power compared to typical manual references.
Key insights
- Retrieval-Augmented Generation (RAG) is critically significant for LLM-driven hardware design, as it effectively mitigates hallucinations and output token limitations, enabling the generation of complete and accurate designs.
- A meticulously curated, domain-specific dataset is essential for fine-tuning LLMs to reliably generate complex hardware, especially when exploring nuanced approximate computing trade-offs.
- Effective prompt template generation is crucial for guiding fine-tuned LLMs towards desired hardware configurations, dramatically improving the success rate of generating top-level design files.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.