GPU-EMI-TELEMETRY

GPU-Accelerated EMI Filtering for High-Speed Automotive Infotainment Streams

GPU-Accelerated EMI Filtering for High-Speed Automotive Telemetry

📌 Overview

Modern electric and autonomous vehicles generate millions of telemetry samples per second. [cite_start]High-power components like traction inverters create broadband switching noise and high-energy electromagnetic interference (EMI)[cite: 24].

[cite_start]This project implements a high-throughput, parallelized GPU processing engine written in CUDA C++ for real-time anomaly detection and "blind" filtering[cite: 11, 26]. [cite_start]It shifts digital signal processing (DSP) workloads from a linear CPU bottleneck to a fully parallel SIMT execution model, achieving a 31.64x speedup over traditional sequential methods[cite: 14, 19].

🚀 Key Features

• [cite_start]Zero-Copy DMA Transport: Utilizes pinned memory (cudaHostAlloc) to bypass CPU staging buffers, reducing PCIe data transport latency by 6.14x[cite: 49, 101, 104].
• [cite_start]$O(1)$ Parallel Thresholding: Implements a custom frequency-gated CUDA kernel that evaluates and suppresses multiple concurrent fault frequencies simultaneously, independently of the number of anomalies[cite: 51, 84].
• [cite_start]In-Place VRAM Optimization: Uses shared memory pointers for forward and inverse FFTs, cutting the active VRAM footprint in half (from ~33.5 MB to 16.77 MB)[cite: 106, 107].

🛠️ Architecture Pipeline

[cite_start]The engine executes a 4-stage edge-computing pipeline[cite: 45]:

1. [cite_start]Data Ingestion: Host CPU ingests raw telemetry arrays (e.g., 4.19M points)[cite: 48].
2. [cite_start]Zero-Copy Transport: Data is pulled directly into the GPU via Direct Memory Access (DMA)[cite: 59].
3. [cite_start]GPU DSP Execution: * Forward Transform using cuFFT (Real-to-Complex)[cite: 61].
   • [cite_start]Parallel Anomaly Suppression via custom SIMT dynamic thresholding kernel[cite: 62].
   • [cite_start]Inverse Transform using cuFFT (Complex-to-Real)[cite: 64].
4. [cite_start]Signal Extraction: Host extracts the normalized base signal utilizing cuBLAS scaling[cite: 65, 66].

📊 Benchmarks & Results

[cite_start]Tested on a 4.19 million point dataset with three simultaneous high-energy mechanical/electrical faults[cite: 116, 117]:

• [cite_start]GPU Latency: 4.451 ms [cite: 122]
• [cite_start]CPU Latency (FFTW3): 140.807 ms [cite: 120]
• [cite_start]True Scientific Speedup: 31.64x [cite: 122]
• [cite_start]Signal Recovery: Complete anomaly deletion while perfectly preserving the ground-truth 60Hz mechanical baseline[cite: 129, 130].

💻 Dependencies

To compile and run this project, you will need:

• [cite_start]NVIDIA CUDA Toolkit (v12.5+) [cite: 118]
• [cite_start]cuFFT Library [cite: 94]
• [cite_start]cuBLAS Library [cite: 95]
• [cite_start]FFTW3 (for CPU baseline benchmarking) [cite: 120]

⚙️ How to Run

1. Generate the synthetic telemetry arrays (creates binary files):

   python generate_telemetry.py