Structural optimization principles for edge AI in motorsport telemetry.

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Engineering decisions in motorsport are constrained by physical limitations like mass and stiffness.
• Deploying large language models (LLMs) at the edge faces similar constraints in memory, bandwidth, latency, and power.
• A systems-engineering approach is needed to bridge these domains for efficient edge AI.

Purpose of the Study:

• To present a systems-engineering framework relating structural lightweighting and post-training quantization (PTQ) of LLMs.
• To demonstrate how second-order sensitivity analysis guides safe material removal in structures and precision reduction in neural networks.
• To evaluate the feasibility of low-precision LLM deployment under motorsport-inspired resource constraints.

Main Methods:

• Developed a framework linking structural topology optimization and LLM PTQ using second-order sensitivity.
• Utilized a Formula Student upright case study for structural analysis and GPTQ/AWQ for LLM evaluation.
• Defined Digital Factor of Safety (DFS) and Intelligence-per-Watt (IPW) metrics for feasibility assessment.

Main Results:

• Sensitivity-guided material removal preserved mechanical integrity in the structural example.
• Sensitivity-aware quantization reduced LLM computational footprint while maintaining usefulness for telemetry inference.
• INT4 weight-only deployment of a 32B LLM reduced memory by 70%, increased throughput by 169%, and lowered power consumption by 44%.

Conclusions:

• The proposed framework enables analysis of low-precision LLM deployment feasibility under strict resource constraints.
• Stiffness-informed digital lightweighting is practically relevant for edge AI in resource-limited environments.
• The methodology provides a reproducible approach for assessing LLM deployment under motorsport-inspired constraints.