Gaming Telemetry: Neuromorphic Data Collector for SNN Training

Overview

High-frequency GPU/CPU telemetry for a workstation (optimized for RTX 5080 under max-settings games). The collector is game-agnostic: it reads NVIDIA NVML + Linux CPU sensors and writes Parquet batches for neuromorphic / SNN training.

Signals map roughly to an artificial “nervous system” for models that need to learn how compute load moves a GPU:

• Excitatory: PCIe floods, power/clock spikes, VRAM allocation jumps
• Inhibitory: thermal / power throttle bitmasks
• State / momentum: fan speed, absolute VRAM usage

There is no game-install verifier, Steam/Proton discovery, or mod scanner. Graphics settings are an operator checklist, not code. New titles only need a new SESSION_LABEL and a play session.

Captured metrics

• Power usage & temperature
• Graphics & memory clocks
• PCIe Rx/Tx throughput
• Performance state & throttle reasons
• Fan speed, VRAM used/total
• Encoder/decoder utilization
• CPU Tctl / CCD temps and package power (hwmon + RAPL energy delta)
• session_label (string; same for every row in a run)

MangoHud (or any overlay) is not recorded. You may still run it yourself for on-screen monitoring; the collector only writes hardware telemetry.

Prerequisites

• OS: Linux (developed on Fedora)
• GPU: NVIDIA with NVML (RTX 50-series preferred)
• Build: Rust / Cargo

Usage

1. Capture a labeled session

Use a dedicated per-session directory so batches from different runs do not overwrite each other.

# Examples: kcd2, re2r, re3r, re_requiem, cp2077, …
TS=$(date +%Y%m%d_%H%M%S)
SESSION_DIR="neuromorphic_data/kcd2_${TS}"
mkdir -p "$SESSION_DIR" && cd "$SESSION_DIR"
SESSION_LABEL=kcd2 cargo run --release --bin gaming-telemetry --manifest-path ../../Cargo.toml
# Ctrl+C to flush, then cd back for the next session
cd ../..

TS=$(date +%Y%m%d_%H%M%S)
SESSION_DIR="neuromorphic_data/re2r_${TS}"
mkdir -p "$SESSION_DIR" && cd "$SESSION_DIR"
SESSION_LABEL=re2r cargo run --release --bin gaming-telemetry --manifest-path ../../Cargo.toml
# Ctrl+C to flush, then cd back for export/query steps
cd ../..

Then:

1. Set the game to the highest graphics settings available.
2. Play the session while the collector runs (default poll: 5 ms, override with POLL_INTERVAL_MS).
3. Ctrl+C to flush the last batch and exit.

Output files (multiple batches per directory):

gpu_telemetry_v2_batch_N.parquet

The export and query examples below reuse the $SESSION_DIR variable from the capture block you ran. If you used a different directory, substitute its name.

2. Export canonical CSV for corinth-canal

Stable 5-column replay schema (unchanged; session_label stays in Parquet):

cargo run --bin export_csv -- "$SESSION_DIR/gpu_telemetry_v2_batch_1.parquet" canonical.csv

Header:

timestamp_ms,gpu_temp_c,gpu_power_w,cpu_tctl_c,cpu_package_power_w

gpu_power_w is power_usage_mw / 1000.0.

3. Optional: DuckDB query helper

cargo run --bin query -- "$SESSION_DIR/gpu_telemetry_v2_batch_1.parquet"

Replay contract

collector -> neuromorphic_data/<session>/gpu_telemetry_v2_batch_N.parquet -> export_csv -> canonical.csv -> corinth-canal/examples/csv_replay

# From the corinth-canal checkout (not this repo root):
cargo run --example csv_replay --manifest-path path/to/corinth-canal/Cargo.toml -- canonical.csv

For multi-title training mixes, group by Parquet session_label (or by folder under neuromorphic_data/).

Operator checklist (not code)

Title	Suggested SESSION_LABEL
Kingdom Come Deliverance 2	kcd2
Resident Evil 2 Remake	re2r
Resident Evil 3 Remake	re3r
Resident Evil Requiem	re_requiem
Cyberpunk 2077	cp2077

Max settings only. No install path is required by this repo.

Design notes

• Collector never walks $HOME, Steam libraries, or Proton prefixes.
• Path redaction helpers remain for error logs / query display only.
• The old Cyberpunk workload verifier direction (PR #6 and residual skeleton/CI) was removed; see issue #20 / Linear RM-174.

License

GPL-3.0. See LICENSE.