rmems/gaming-telemetry

🎮 Gaming Telemetry

High-frequency bare-metal GPU/CPU telemetry captured while playing games at maximum settings — DLSS 4.0, path tracing, the works — on an NVIDIA RTX 5080 workstation. Collected for neuromorphic / spiking-neural-network research: a game session turns out to be a far richer stress instrument than crypto mining or sync-node workloads, with genuine idle → ramp → sustained-load → recovery dynamics, real wall-clock timestamps, and the NVML throttle bitmask.

This is the canonical multi-title hardware corpus for the Spikenaut / Limen-Neural ecosystem. One config per capture session; every config shares the same collector schema, Parquet only.

from datasets import load_dataset

re4 = load_dataset("rmems/gaming-telemetry", "re4", split="train")

📊 Sessions

Config	Title	Rows	Span	Cadence (mean)	Notes
re4	Resident Evil 4 Remake (path tracing)	95,893	90.1 min	56.4 ms	9.3% of samples at SW power cap (0x4), peak 364.9 W, max 66 °C, VRAM to 15.9 GB

Future sessions (Cyberpunk 2077, KCD2, …) land as new configs — the collector is game-agnostic; a new title only needs a SESSION_LABEL and a play session.

🔍 Schema (system_telemetry_v1 + session_label)

Column	Type	Unit / meaning
timestamp_ms	int64	wall clock, ms since Unix epoch
power_usage_mw	uint32	GPU board power, milliwatts
temperature_c	uint32	GPU core temperature
graphics_clock_mhz, memory_clock_mhz	uint32	clocks
pcie_rx_kbps, pcie_tx_kbps	uint32	PCIe throughput
pstate	uint32	NVML performance state (P0=0 …)
throttle_reasons_bitmask	uint64	nvmlClocksThrottleReasons bits (0x4 = SW power cap, 0x20/0x40 = SW/HW thermal, 0x80 = HW power brake)
fan_speed_perc	uint32	fan duty
memory_used_mb, memory_total_mb	uint64	VRAM
encoder_util_perc, decoder_util_perc	uint32	NVENC/NVDEC utilization
cpu_tctl_c, cpu_ccd1_c, cpu_ccd2_c	float32	CPU temperatures (hwmon)
cpu_package_power_w	float32	CPU package power (RAPL energy delta)
session_label	string	capture session id, constant per config

Missing readings are null, never 0.

📈 Provenance

rmems/gaming-telemetry (GitHub)   Rust collector: NVML + hwmon/RAPL → Parquet batches
        ↓
rmems/gaming-telemetry (this dataset)   canonical per-session configs
        ↓
rmems/SEMM-Latent-Telemetry       SAAQ / SEMM routing research (RE4 was its hardware drive)
rmems/Spikenaut-SNN-Telemetry     supervisor-trajectory dataset (state_telemetry backfill candidate)

re4 normalization, for the record. The seed capture was first published as 48 raw batch files in SEMM-Latent-Telemetry/origin_hardware_baselines/resident_evil_4/ (system_telemetry_v1_batch_1..48.parquet), which remain the untouched origin artifact. This dataset's re4/train-00000.parquet is those 48 batches concatenated, sorted by timestamp_ms (lexicographic batch globbing would interleave time), with a constant session_label = "re4" column appended to match the collector's v2 convention. Row count (95,893), null counts, and per-column sums were verified identical to the raw batches before publishing; no value was altered.

The capture predates the collector's v2 schema, so v2-only fields beyond session_label are absent here; future sessions will carry the current collector schema and a documented POLL_INTERVAL_MS.

⚠️ Honest caveats

• Single machine, single session so far. One RTX 5080 workstation, one ~90-minute RE4 run. No cross-GPU generality is claimed.
• The effective cadence (~56 ms mean) is the observed spacing of this capture, not a declared collector setting for it.
• The GPU never left its comfort zone: max 66 °C, no thermal or HW-slowdown throttle bits fired — the only non-zero mask is SW power cap. Good "healthy under load" baseline; not an incident corpus.
• Gameplay content is not recorded — no frames, inputs, audio, or overlay data. Hardware counters only.

📜 Citation

@dataset{gaming_telemetry,
  author={Montoya Cardenas, Raul},
  title={Gaming Telemetry: bare-metal GPU/CPU traces under max-settings gameplay},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/rmems/gaming-telemetry}
}

⚖️ License

MIT OR Apache-2.0 — dual-licensed; use whichever fits your project.

🙏 Acknowledgments

• Capcom's RE Engine team — unknowingly the best GPU stress instrument in this lab
• Claude Fable 5 (Anthropic) — dataset normalization, validation, and this card, in Claude Code under human review

Built by Raul Montoya Cardenas — WGU AI Engineering