Presentation
NPU-Based Energy-Efficient LLM Inference for Semiconductor Equipment Control Code Generation
DescriptionIndustrial Algorithmic Pattern Generator (ALPG) code generation
with large language models (LLMs) has shown strong results under
high-cost configurations using proprietary models and GPU-based
pipelines. However, such setups limit practical deployment in on-
premise semiconductor environments. In this work, we ask how
far a cost-constrained pipeline can approach industrial viability.
We deploy an FP8-quantized open-source Llama 3.3 model on an
emerging low-power NPU accelerator (Renegade) and evaluate
domain-specific code generation. Compared to an NVIDIA A100
GPU, the NPU achieves comparable throughput (80–110 tokens/s)
while reducing steady-state power by 50–70%, yielding 2–3× higher
tokens-per-watt efficiency. Structural validity improves from 40% to
60–65% with lightweight rule-based correction. These results high-
light the trade-off between energy efficiency and domain accuracy,
suggesting that energy-efficient NPUs can approach practical indus-
trial deployment without relying on high-cost model and hardware
configurations.
with large language models (LLMs) has shown strong results under
high-cost configurations using proprietary models and GPU-based
pipelines. However, such setups limit practical deployment in on-
premise semiconductor environments. In this work, we ask how
far a cost-constrained pipeline can approach industrial viability.
We deploy an FP8-quantized open-source Llama 3.3 model on an
emerging low-power NPU accelerator (Renegade) and evaluate
domain-specific code generation. Compared to an NVIDIA A100
GPU, the NPU achieves comparable throughput (80–110 tokens/s)
while reducing steady-state power by 50–70%, yielding 2–3× higher
tokens-per-watt efficiency. Structural validity improves from 40% to
60–65% with lightweight rule-based correction. These results high-
light the trade-off between energy efficiency and domain accuracy,
suggesting that energy-efficient NPUs can approach practical indus-
trial deployment without relying on high-cost model and hardware
configurations.
Event Type
Work in Progress
TimeMonday, July 276:14pm - 6:15pm PDT
LocationExhibit Hall