Enforce GPU performance discipline with ncu, nsys, and nvidia-smi profiling.
npx skills add https://github.com/ianrgraham/gale --skill gale-gpu-perf
Please help me install this Agent Skill. Skill: gale-gpu-perf Source: https://github.com/ianrgraham/gale/tree/main/.claude/skills/gale-gpu-perf Command: npx skills add https://github.com/ianrgraham/gale --skill gale-gpu-perf
This Skill prevents GPU changes in gale-gpu from becoming correct-but-slow by enforcing device residency, persistent handles, and mandatory profiling before the work is considered done.
Use this skill whenever you modify GPU code in gale-gpu and report the profiling results before declaring the change complete.
High-intent search queries and answers about installing and using this skill.
Profiling GPU code with nsys, ncu, and nvidia-smi validates kernel performance by measuring actual wall-clock behavior. This profiling gate ensures CUDA modifications are judged by real execution metrics rather than assumptions, preventing correct-but-slow code.
Device residency in GPU simulation paths means keeping simulation state on the GPU and driving time stepping without host synchronization. It is required for production-ready CUDA kernels and integrator paths that must remain fast and fully device-resident during execution.
Profile modified CUDA launch wrappers by running nsys, ncu, and nvidia-smi to capture real performance metrics. You must report these profiling results and verify the absence of per-step synchronization before declaring the kernel change complete.
Yes, this GPU performance discipline applies to Rust-based CUDA integrators and host launch wrappers. It enforces persistent handles and device residency while requiring ncu and nsys validation to ensure Rust GPU code remains fast and sweep-ready.
GPU code runs slowly during time stepping if it uses one-shot kernel loads or per-step host-device synchronization. Enforcing device residency and persistent handles eliminates these GPU utilization killers, ensuring simulation state stays on the device.
BagelHoleConfigure NVIDIA GPU servers with drivers, CUDA, and MIG partitioning.
NVIDIAAnalyze Nsight Compute .ncu-rep reports to classify CUDA kernel bottlenecks.
igamenovoerConfigure CUDA build environments for Pixi projects with CMake and nvcc.
rsmdtDiagnose performance issues by profiling CPU, memory, and I/O metrics.
