Gaming Telemetry: Neuromorphic Data Collector for SNN Training

Overview

This high-performance Rust daemon is designed to capture high-fidelity GPU telemetry data from demanding gaming workloads. Specifically optimized for systems running Resident Evil 4 and Cyberpunk 2077 with Path Tracing and DLSS 4.0, it provides the rich, high-frequency time-series data required to train Spiking Neural Networks (SNNs) and Liquid State Machines.

The collector identifies "excitatory" spikes (e.g., PCIe bus floods during asset loading) and "inhibitory" signals (e.g., thermal throttling or power caps), mimicking the dynamics of biological neural systems.

Key Features

• Ultra-Low Latency Polling: Captures metrics at 5-millisecond intervals using the NVIDIA Management Library (NVML).
• Asynchronous I/O: To prevent performance drops during heavy gaming (Path Tracing), data is buffered in memory and written to versioned Parquet files (gpu_telemetry_v1_batch_N.parquet) asynchronously using tokio and polars.
• DuckDB Integration: Includes a built-in query utility for instant analysis of the captured Parquet batches.
• Rich Metric Suite: Captures complex hardware states beyond simple temperature and power.

Captured Metrics

The telemetry captures a blend of fast-moving transients and slow-moving momentum metrics:

• PCIe Rx/Tx Throughput: Detects data floods from the CPU/Memory (e.g., BVH structure updates for Path Tracing).
• Power Usage & Temperature: High-frequency transients.
• Graphics & Memory Clocks: Tracking the "firing rate" of the silicon.
• Throttle Reasons: Captures bitmasks for Power, Thermal, and Sync limits (Inhibitory signals).
• Fan Speed (RPM): A slow-moving physical momentum metric.
• VRAM Utilization: Tracks spatial memory pressure and allocation spikes.

Prerequisites

• OS: Fedora 43 Linux
• GPU: NVIDIA (Optimized for RTX 50-series, compatible with others)
• Drivers: Proprietary NVIDIA drivers with NVML support.
• Build Tools: Rust (Cargo)

Usage

1. Start the Telemetry Daemon

Run the daemon in release mode to ensure minimal overhead and maximum timing accuracy.

cargo run --release --bin gaming-telemetry

The daemon continuously polls telemetry and writes versioned batches as:

gpu_telemetry_v1_batch_N.parquet

CPU package power is recorded from the CpuMonitor time-delta energy-counter path.

2. Export Canonical CSV for corinth-canal

Convert one v1 Parquet batch into the stable 5-column replay schema:

cargo run --bin export_csv gpu_telemetry_v1_batch_1.parquet canonical.csv

Canonical CSV header (exact order):

timestamp_ms,gpu_temp_c,gpu_power_w,cpu_tctl_c,cpu_package_power_w

gpu_power_w is exported as power_usage_mw / 1000.0. CPU columns come from recorded parquet columns.

3. Optional: Analyze Data with DuckDB

Use the query utility for ad-hoc analysis:

cargo run --bin query gpu_telemetry_v1_batch_1.parquet

Replay Contract

One-way flow:

collector -> gpu_telemetry_v1_batch_N.parquet -> export_csv -> canonical.csv -> corinth-canal/examples/csv_replay

Consumer command in corinth-canal:

cargo run --example csv_replay canonical.csv

Architecture for SNNs

The data collected is structured to be directly useful for Neuromorphic computing:

• Excitatory Inputs: PCIe throughput and VRAM allocation rate.
• Firing Rates: Clock speeds and Power transients.
• Inhibitory Inputs: Thermal/Power throttling bitmasks.
• State/Momentum: Fan speeds and absolute VRAM usage.

License

GPL-3.0 License. See LICENSE for details.