derek.holloway
7502bf64a4
1.5 KiB
status, author, start_date, pr, implements, supersedes
| status | author | start_date | pr | implements | supersedes |
|---|---|---|---|---|---|
| accepted | aelwazir | 2026-04-17 | https://example.com/pr/0 |
RFC-0000: Canonical example (non-binding)
Summary
This file is a worked example demonstrating the RFC format. It is not a live proposal and no implementation follows from it.
Motivation
Contributors benefit from seeing a completed RFC before writing their first one.
Design
(Illustrative; not implemented.) Add a "thermal-throttle detection" signal
to the latency bottleneck classifier that reads temperature counters
when available.
Public-facing change
perfxpert analyze would emit a new recommendation category
thermal_throttling when TEMPERATURE_EDGE > 95 C correlates with
kernel slowdown.
Internal change
- New tool
tools/thermal.py:classify_throttle - New knowledge entry
thermal_thresholds.yaml - New bottleneck sub-type in
bottleneck_types.yaml
Backward compatibility
Non-breaking. If thermal counters absent, classifier skips silently.
Alternatives considered
(1) OS-level thermal probe — rejected: requires root. (2) Ignore thermal entirely — rejected: leaves a real class of slowdowns invisible.
Unresolved questions
- What per-arch threshold? Differs MI300X vs MI350X.
Test plan
Tool unit tests, integration test with a thermal-throttled fixture, proven-optimization entry with fixture-pair.
Migration
n/a (additive)
Prior art
AMD SMI / ROCm SMI thermal telemetry and existing ROCm monitoring flows.
Approval checklist
- Example (non-binding)