Clemson University
Domain-Specific Design on FPGA and NPU for scientific computing and On-Device LLM Inference
Abstract
dc:description.abstractHigh-performance computing (HPC) is changing rapidly as scientific simulations and large language model (LLM) workloads push the need for higher performance under tight power and memory constraints. Conventional platforms such as CPUs and GPUs accelerate computation through instruction-driven parallelism, relying on multithreading, SIMD, and SIMT execution, but increasingly encounter scalability limits imposed by the power and memory walls. In contrast, Field-Programmable Gate Arrays (FPGAs) and Neural Processing Units (NPUs) offer a high-efficiency alternative through dataflow-oriented architectures that exploit deep pipelining and customized memory hierarchies to reduce data movement. However, the performance potential of these spatial accelerators remains largely unrealized when traditional, control-flow-centric algorithms are directly mapped onto them. This dissertation addresses this gap by developing domain-specific algorithm and dataflow designs tailored for FPGAs and NPUs, demonstrating that hardware–software co-design is essential for achieving high performance and energy efficiency on modern accelerators. Two representative and challenging applications are studied: electromagnetic simulation using the Finite-Difference Time-Domain (FDTD) method and on-device inference for large language models. FDTD simulations are critical for the design of photonic integrated circuits but are computationally intensive, while on-device LLM inference requires low latency and low power consumption. For FDTD, this work introduces a time-pipelined computation that significantly reduces data movement and enables scalable execution across FPGA networks and reconfigurable accelerators, substantially shortening photonic design cycles. For large language models, it demonstrates how reorganizing computation and dataflow allows NPUs to process long sequences efficiently, achieving lower latency and energy consumption than existing approaches.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (PhD)
- Level thesis:degree_level
- Dissertation
- Discipline thesis:degree_discipline
- Electrical and Computer Engineering
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Yu, Miaoxiang
- Contributors dc:contributor
-
- Dr.Tao Wei
- Dr. Fatemeh Afghah
- Dr. Rong Ge
- Dr. Judson Ryckman
Subjects
dc:subject × 6Identifiers
dc:identifier.*- Repository record dc:identifier
- https://open.clemson.edu/all_dissertations/4205