gpu: lock beast's memory clock at 810 MHz now that it mines

Measured 2026-09-09 with both 5090s at 400 W: 555 -> 567 and 563 -> 573
MH/s, SM clock 2250 -> 2370 MHz. Applied to beast with the gpu role.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CBgs2nSi4H2mdh8kD8vMX5
This commit is contained in:
Rob Thijssen
2026-09-09 09:27:02 +03:00
co-authored by Claude Fable 5.1
parent c8aef85e24
commit 921f468b2f
+5 -3
View File
@@ -67,9 +67,11 @@ GPU_POWER_LIMITS="${GPU_POWER_LIMITS-beast.hanzalova.internal=300 benjy.hanzalov
# SMs inside the same power cap: locking it at 810 MHz measured +13% on the
# 4090 at 250 W (410 -> 463 MH/s) and +2% on the 3060 at 130 W. 405 MHz drops
# the card into a low-power state and must not be used. Hosts serving
# inference (beast) are NOT listed: LLM inference is memory-bound. A host with
# no entry gets its memory clock reset to the driver default.
GPU_MEMORY_CLOCKS="${GPU_MEMORY_CLOCKS-benjy.hanzalova.internal=810 quadbrat.hanzalova.internal=810}"
# beast joined 2026-09-09 once inference was switched off there (+2% on each
# 5090 at 400 W, 555 -> 567 MH/s, SM clock 2250 -> 2370 MHz); a host serving
# LLM inference must NOT be listed, inference is memory-bound. A host with no
# entry gets its memory clock reset to the driver default.
GPU_MEMORY_CLOCKS="${GPU_MEMORY_CLOCKS-benjy.hanzalova.internal=810 quadbrat.hanzalova.internal=810 beast.hanzalova.internal=810}"
# Vanity name served by the edge proxy. Dot-free short name: the step@ renewal
# unit is a template instance and %i cannot carry dots (internal-tls.md §1).
EDGE_NAME="${EDGE_NAME:-quantus}"