With the increasingly prominent characteristics of data-intensive and AI-driven scientific paradigms, computing power has become a crucial pillar of research activities. This study aims to examine the knowledge innovation effects of university supercomputing development by theoretically proposing two mechanisms: the efficiency effect (including technical and nontechnical factors) and the scale effect. Empirically, we match the extracted scientific publications of universities via Web of Science with supercomputer information from the top 500 rankings, constructing a panel dataset of 110 distinguished universities in China. Within the causal inference framework, the spatial difference-in-differences method (spatial DID) is employed to assess the impact of university computing power "upgrades" on knowledge innovation. The research findings include: (1) Overall, supercomputing construction stimulates knowledge innovation in universities, primarily manifested in the increase in the number of general papers and highly cited papers cataloged by the Science Citation Index (SCI). (2) Knowledge innovation effects have a lag period of approximately one--four years and may have a negative impact on innovation in geographically and economically adjacent universities. (3) Supercomputing construction mainly promotes university knowledge innovation by improving innovation efficiency (efficiency effect), accounting for 85.0% [approximately] 96.5% of the total effect. Among these factors, the proportion of nontechnical factors is at most 38.0%. In contrast, the scale effect accounts for a maximum of 15.0%, which is achieved mainly through an increase in the scale of research personnel. (4) There is significant interuniversity heterogeneity in the knowledge innovation effects of supercomputing, with Tsinghua University, Shanghai Jiao Tong University, and Jilin University showing the most significant effects. Additionally, we provide a series of potential optimization utility lists for universities, which, together with benchmark and mechanism tests, constitute a complete policy sandbox.