What problem does it solve? DeepStream video-analytics pipelines are typically configured by guesswork, leaving GPU decode, compute, or memory-bandwidth bottlenecks undiagnosed. This Skill replaces trial-and-error tuning with a measured 6-stage workflow that identifies the actual bottleneck and derives every config knob from two measured numbers: inference plateau batch and hardware ceilings. ## Core Features & Use Cases - Stage 0 perf-correct presets: Pre-applies INT8/FP16 precision, NVMM zero-copy memory, model-dimension streammux, and fakesink defaults before a pipeline is even generated. - Measurement-driven profiling: Runs an inference-only micro-benchmark sweep, queries nvidia-smi for NVDEC/compute/bandwidth ceilings, and captures end-to-end traces with nsys profile and nsys stats. - Bottleneck classification and capacity report: Classifies the pipeline as DECODE_BOUND, COMPUTE_BOUND, MEMORY_BW_BOUND, TRACKER_BOUND, or SYNC_BOUND, then computes max-streams capacity with bottleneck-specific remediation via the capacity_report.py script. - Use Case: An engineer asks how many 1080p30 H.264 streams a GPU can handle for ResNet18 detection. The Skill runs the micro-benchmark, classifies the bound type, and reports a concrete max-streams number plus which hardware upgrade actually helps. ## Quick Start Ask the agent to profile your DeepStream pipeline or ask how many streams your GPU can handle for your model at your target FPS.
