Large Reasoning Models (LRMs) are becoming increasingly popular as they offer advanced capabilities in logical inference, mathematical reasoning, and knowledge synthesis, even beyond those of standard language models. However, their complex training workflows present significant challenges in reproducibility, efficiency, and system-level optimization. This paper introduces HPC-R1, a comprehensive characterization of LRM training on the NERSC Perlmutter supercomputer, representing behavior on a Top500-ranked system. We analyze all major stages, including supervised fine-tuning (SFT), Group Relative Policy Optimization (GRPO)-based reinforcement learning (RL), autoregressive generation, and distillation using customized state-of-the-art frameworks. Our detailed performance analysis reveals key system inefficiencies and scaling behaviors. Through our in-depth analysis, we present 19 key observations across all stages, including 4 for SFT, 7 for GRPO-based RL, 6 for generation, and 2 for distillation. Based on these findings, we present several key recommendations to guide future HPC-AI system design.

High resolution fractional vegetation cover (HR-FVC) is important for many applications, including precision agriculture, forestry, and conservation. For land managers, HR-FVC is most useful when the data can be produced quickly with minimal effort. In this study, we perform data fusion of RGB drone data and multispectral cubesat data for synthetic daily HR-FVC estimation. First, binary classification of 10cm resolution drone data was used to identify vegetation. An AdaBoost model (Accuracy = 0.868, F1-score = 0.840) was selected for further analysis. HR-FVC training data was then produced from drone vegetation maps by calculating the FVC in a 3m pixel – Planet SuperDove resolution, resulting in 238,270 training points. A random forest regression model was used to predict HR-FVC from Planet SuperDove data. The final model’s performance is comparable to similar studies (R2= 0.720, RMSE = 0.213), suggesting the methodology could be viable for applications requiring daily HR-FVC datasets.