twitch-highlight-detector/monitoring_report.md

GPU/CPU Monitoring Report - Twitch Highlight Detector

System Information

• GPU: NVIDIA GeForce RTX 3050 (8192 MiB)
• Driver: 580.126.09
• Device: cuda (CUDA requested and available)

Execution Summary

• Total Runtime: ~10.5 seconds
• Process Completed: Successfully
• Highlights Found: 1 (4819s - 4833s, duration: 14s)

GPU Utilization Analysis

Peak GPU Usage

• Single Peak: 100% GPU SM utilization (1 second only)
• Location: During RMS calculation phase
• Memory Usage: 0-4 MiB (negligible)

Average GPU Utilization

• Overall Average: 3.23%
• During Processing: ~4% (excluding idle periods)
• Memory Utilization: ~1% (4 MiB / 8192 MiB)

Timeline Breakdown

1. Chat Analysis: < 0.1s (CPU bound)
2. FFmpeg Audio Extraction: 8.5s (CPU bound - FFmpeg threads)
3. Audio Decode: 9.1s (CPU bound - soundfile library)
4. CPU->GPU Transfer: 1.08s (PCIe transfer)
5. GPU Processing:
   • Window creation: 0.00s (GPU)
   • RMS calculation: 0.12s (GPU - 100% spike)
   • Peak detection: 0.00s (GPU)

CPU vs GPU Usage Breakdown

CPU-Bound Operations (90%+ of runtime)

1. FFmpeg audio extraction (8.5s)

   • Process: ffmpeg
   • Type: Video/audio decoding
   • GPU usage: 0%

2. Soundfile audio decoding (9.1s overlap)

   • Process: Python soundfile
   • Type: WAV decoding
   • GPU usage: 0%

3. Chat JSON parsing (< 0.5s)

   • Process: Python json module
   • Type: File I/O + parsing
   • GPU usage: 0%

GPU-Bound Operations (< 1% of runtime)

1. Audio tensor operations (0.12s total)

   • Process: PyTorch CUDA kernels
   • Type: RMS calculation, window creation
   • GPU usage: 100% (brief spike)
   • Memory: Minimal tensor storage

2. GPU Memory allocation

   • Audio tensor: ~1.2 GB (308M samples × 4 bytes)
   • Chat tensor: < 1 MB
   • Calculation buffers: < 100 MB

Conclusion

FAIL: GPU not utilized

Reason: Despite the code successfully using PyTorch CUDA for tensor operations, GPU utilization is minimal because:

1. Bottleneck is CPU-bound operations:

   • FFmpeg audio extraction (8.5s) - 0% GPU
   • Soundfile WAV decoding (9.1s) - 0% GPU
   • These operations cannot use GPU without CUDA-accelerated libraries

2. GPU processing is trivial:

   • Only 0.12s of actual CUDA kernel execution
   • Operations are too simple to saturate GPU
   • Memory bandwidth underutilized

3. Architecture mismatch:

   • Audio processing on GPU is efficient for large batches
   • Single-file processing doesn't provide enough parallelism
   • RTX 3050 designed for larger workloads

Recommendations

To actually utilize GPU:

1. Use GPU-accelerated audio decoding:

   • Replace FFmpeg with NVIDIA NVDEC
   • Use torchaudio with CUDA backend
   • Implement custom CUDA audio kernels

2. Batch processing:

   • Process multiple videos simultaneously
   • Accumulate audio batches for GPU
   • Increase tensor operation complexity

3. Alternative: Accept CPU-bound nature:

   • Current implementation is already optimal for single file
   • GPU overhead may exceed benefits for small workloads
   • Consider multi-threaded CPU processing instead

Metrics Summary

• GPU utilization: 3.23% average (FAIL - below 10% threshold)
• CPU usage: High during FFmpeg/soundfile phases
• Memory usage: 4 MiB GPU / 347 MB system
• Process efficiency: 1 highlight / 10.5 seconds