config-multisignal.toml

# unifand configuration — multi-signal fusion, tuned starting point
#
# Fans follow the highest duty any enabled signal's curve commands
# (raise-only: a signal can raise duty above what die temp asks for, never
# lower it). This file is a STARTING POINT, not a drop-in: the curve knees
# below were measured on one RTX 5090 FE (575 W enforced limit, 90 °C
# thermal limit) and every card needs its own tuning pass. The gpu_power
# curve ships in percent_tdp, which is portable — percent of whatever limit
# your card enforces; the commented watts alternative is NOT portable and
# must be retuned per card. See README "Multi-signal fusion".

poll_interval_secs = 5
channels = [1, 2]
fallback_duty = 60
cpu_offset = 10.0

# NOTE: hysteresis_c is honored in legacy (signals-disabled) mode only.
# With fusion enabled it is inert — fusion decides in duty space, so change
# suppression degenerates to min_duty_delta.
hysteresis_c = 2.0
min_duty_delta = 5
max_step_per_tick = 10

# Top-level curve: still required, and doubles as the die-temp curve for
# the gpu_temp/cpu_temp signals below (they fall back to it when they have
# no curve of their own).
[[curve]]
temp = 35.0
duty = 30
[[curve]]
temp = 55.0
duty = 40
[[curve]]
temp = 68.0
duty = 65
[[curve]]
temp = 75.0
duty = 85
[[curve]]
temp = 80.0
duty = 100

[metrics]
listen = "0.0.0.0:9877"

[signals]
enabled = true

# Die temperatures: same behavior as the legacy loop (both fall back to the
# top-level [[curve]]), now competing as candidates alongside the signals
# below.
[signals.gpu_temp]
enabled = true

[signals.cpu_temp]
enabled = true
offset_c = 10.0

# Thermal margin: headroom in °C to the card's own thermal limit
# (limit - die temp). INVERTED curve — small headroom means high duty.
# Self-calibrating: this exact curve is correct on any card, because it is
# denominated in distance-to-the-limit rather than absolute degrees.
# Knees from measured data: 5 °C headroom under sustained load, 39-55 °C
# at idle.
[signals.thermal_margin]
enabled = true
[[signals.thermal_margin.curve]]
value = 5
duty = 100
[[signals.thermal_margin.curve]]
value = 10
duty = 85
[[signals.thermal_margin.curve]]
value = 20
duty = 60
[[signals.thermal_margin.curve]]
value = 35
duty = 30

# GPU power draw: the leading signal — on load release, power sheds ~96% of
# its range in ~20 s while die temp has shed about half and keeps decaying
# for 40+ s. The asymmetric filter (rise_alpha/fall_alpha defaults 0.5/0.1)
# tracks load onset fast and releases slowly so the fans keep clearing the
# heatsink after power drops.
#
# percent_tdp is the PORTABLE unit: curve X is percent of the card's
# enforced power limit, so the same knees hold across cards. Measured on a
# 575 W limit: idle ~3%, sustained load ~98%.
[signals.gpu_power]
enabled = true
unit = "percent_tdp"
[[signals.gpu_power.curve]]
value = 25
duty = 30
[[signals.gpu_power.curve]]
value = 50
duty = 55
[[signals.gpu_power.curve]]
value = 75
duty = 75
[[signals.gpu_power.curve]]
value = 95
duty = 100

# Absolute-watts alternative, tuned for a 575 W card (RTX 5090 FE) —
# requires retuning for any other card. To use it: set unit = "watts"
# above, delete the percent_tdp curve, and uncomment:
# [[signals.gpu_power.curve]]
# value = 120
# duty = 30
# [[signals.gpu_power.curve]]
# value = 300
# duty = 60
# [[signals.gpu_power.curve]]
# value = 450
# duty = 85
# [[signals.gpu_power.curve]]
# value = 560
# duty = 100

# Throttle duty floor: while the GPU reports any of the listed reasons,
# applied duty is raised to at least floor_duty, held for hold_secs after
# the last assertion (anti-flap). A safety net, not a routine path — no
# thermal throttle bit was ever observed asserting on the measured card,
# even at 84 °C with 5 °C of headroom. sw_thermal (SW_THERMAL_SLOWDOWN)
# fires on GPU *or memory* over-temp, which makes this floor the only
# memory-overtemp response reachable on cards that don't expose a memory
# temperature (see mem_temp below). Do NOT add "sw_power_cap": it asserts
# continuously under any sustained load (measured) and would pin fans at
# the floor for every job.
[signals.throttle]
enabled = true
floor_duty = 85
hold_secs = 30
reasons = ["sw_thermal", "hw_thermal", "hw_power_brake"]

# Memory junction temperature: disabled because NVML reports it
# NotSupported on the measured card (RTX 5090 — nvidia-smi shows N/A too),
# as on most consumer silicon. Only enable it where
# unifand_signal_value{signal="mem_temp"} actually reports; until then the
# throttle floor above is the memory-overtemp path.
[signals.mem_temp]
enabled = false
# [[signals.mem_temp.curve]]
# temp = 80.0
# duty = 50