Noticed a high error rate and significant slowdown (50x slower than before) in Cursor’s ‘code apply process’. Observed high GPU usage (90% under Copy), minimal VRAM and CUDA involvement, and loud fan noise. Restarting the app and system had no effect. Suspect inefficient use of local resources by the small local model.
Issue
Recently, I’ve been facing significant performance issues with Cursor that I believe warrant further clarification and potential solutions.
Cursor seems to use a small local language model to execute this feature, parsing the output from an LLM to the file being edited. However, I am not entirely sure if this model runs locally or if there is another underlying process involved.
At first, I began to notice a higher-than-usual error rate from the small local model that Cursor uses for interpreting and parsing LLM output. This model plays a crucial role by processing the output from the primary LLM and applying code changes. Following this initial issue, the code application process itself started to slow down dramatically, running nearly ten times slower than what I’ve previously experienced, which has greatly affected my workflow.
Soon after, I observed that my computer’s fan began to spin loudly whenever the ‘apply code’ phase was running, suggesting a considerable increase in resource consumption. To understand the cause, I checked the Task Manager and saw that GPU usage spiked significantly, reaching up to 88% under Copy processing. Surprisingly, despite this high usage, VRAM and CUDA processing remained minimal, which pointed to an inefficient utilization of the available GPU resources. This pattern raised questions about the underlying implementation of the local model and how it manages hardware resources during these operations.
Troubleshooting Attempts
Application Restart: I restarted the Cursor application, but this did not resolve the problem.
System Reboot: Restarting the computer also had no impact on the issue.
Seeking Clarification
Expected Behavior: Is it normal for Cursor’s local language model to exhibit such high GPU usage with minimal VRAM and CUDA involvement?
Potential Issues: Could this be a bug or an unintended effect introduced in recent updates?
Recommendations: Are there any known optimizations or solutions to prevent this high GPU load and improve overall performance?
Given that Cursor is a paid product, users should not need to deal with these kinds of development challenges. Ensuring an efficient and reliable user experience is crucial. Any feedback or acknowledgment from the team would be appreciated.
Maybe an extension is using resources. Open the command palette and search for Developer: Open Process Explorer to see if you can identify it. Also, try starting in disable-extensions mode to check if the issue persists.
As the others, I experienced this with a growing codebase. I don’t know if there is anything to be done here but with more files in my workspace, also the wait always increased (as well as the errors/failures to apply changes).
Sorry for the delay in providing a follow-up. After disabling all extensions (I had only the very basic ones) the issue went away. I have also updated Cursor since, so I can’t pinpoint the root cause accurately.