The scaling laws have become the de facto guidelines for designing large language models (LLMs), but they were studied under the assumption of unlimited computing resources for both training and inference. As LLMs are increasingly used as personalized intelligent assistants, their customization (i.e., learning through fine-tuning) and deployment onto resource-constrained edge devices will become more and more prevalent. An urging but open question is how a resource-constrained computing environment would affect the design choices for a personalized LLM. We study this problem empirically in this work. In particular, we consider the tradeoffs among a number of key design factors and their intertwined impacts on learning efficiency and accuracy. The factors include the learning methods for LLM customization, the amount of personalized data used for learning customization, the types and sizes of LLMs, the compression methods of LLMs, the amount of time afforded to learn, and the difficulty levels of the target use cases. Through extensive experimentation and benchmarking, we draw a number of surprisingly insightful guidelines for deploying LLMs onto resource-constrained devices. For example, an optimal choice between parameter learning and RAG may vary depending on the difficulty of the downstream task, the longer fine-tuning time does not necessarily help the model, and a compressed LLM may be a better choice than an uncompressed LLM to learn from limited personalized data.

Compute-in-memory (CIM) accelerators using non-volatile memory (NVM) devices offer promising solutions for energy-efficient and low-latency Deep Neural Network (DNN) inference execution. However, practical deployment is often hindered by the challenge of dealing with the massive amount of model weight parameters impacted by the inherent device variations within non-volatile computing-in-memory (NVCIM) accelerators. This issue significantly offsets their advantages by increasing training overhead, the time needed for mapping weights to device states, energy consumption, and diminishing inference accuracy. To mitigate these challenges, we propose the "Tiny Shared Block (TSB)" method, which integrates a small shared 1x1 convolution block into the DNN architecture. This block is designed to stabilize feature processing across the network, effectively reducing the impact of device variation. Extensive experimental results show that TSB achieves over 20x inference accuracy gap improvement, over 5x training speedup, and weights-to-device mapping cost reduction while requiring less than 0.4% of the original weights to be write-verified during programming, when compared with state-of-the-art baseline solutions. Our approach provides a practical and efficient solution for deploying robust DNN models on NVCIM accelerators, making it a valuable contribution to the field of energy-efficient AI hardware.