LLM Quantization for Edge Hardware

Research Topic & Context

Quantization strategies for deploying large language models on integer-only edge hardware

About this draft

This 85-page extended draft surveys the engineering trade-offs of running large language models on integer-only edge hardware. It compares post-training quantization against quantization-aware training, and works through the practical INT8/INT4 decisions — where accuracy degrades, what memory budget each choice implies, and how latency behaves on real accelerator constraints.

The draft demonstrates CiteLyra's long-form technical writing: an IEEE-style document with a coherent argument sustained across 85 pages, grounded in a bibliography that includes indexed arXiv preprints and peer-reviewed venues. Required metadata, DOI records, and source URLs were checked during generation; readers should still open and assess the sources themselves.

What the 85-page draft covers

• Introduction and problem statement
• Background: transformer inference on integer-only hardware
• Survey of post-training quantization methods
• Quantization-aware training approaches
• INT8 vs INT4: latency, memory, and accuracy trade-offs
• Evaluation methodology and benchmark results discussion
• Conclusion and future work
• IEEE reference list

Create a draft like this

This document was produced with the AI Thesis Writer. Set your topic and constraints, and CiteLyra searches 200M+ real publications, builds a verified bibliography, and drafts a structured document you finish yourself.

More example drafts

• Microfinance and Rural Household Resilience — Standard Draft, 64 pages, APA style
• Supply Chain Resilience After Global Disruptions — Research Paper, 45 pages, Harvard style
• Constitutional Governance of AI Systems — Long-form Research Draft, 119 pages, Chicago style