Introduction

  • The 12V-2x6 connectors are notorious for their flammability.
  • The ASUS ROG ASTRAL implementation of RTX 5090 monitors voltage and current
    flowing through individual power pairs.

  This project provides a native Linux implementation of the current monitoring
  functionality of ASUS ROG ASTRAL devices. The support is currently limited to

    ROG-ASTRAL-RTX5090-O32G
    ROG-ASTRAL-RTX5090-O32G-WHITE
    ROG- STRIX-RTX5090-032G-BTF

  as the authors and contributors do not have access to other devices.

  • The bin/susm binary implements a simple monitor.
  • The bin/susd binary implements a simple daemon that attempts to reduce the
    power draw when it detects overload on any of the wires.

  Both utilities require the i2c-dev kernel module to be loaded.

    modprobe i2c-dev

  The daemon calculates the reduced power. If possible, it tries to use the
  standard limiting facilities. Otherwise it declares a state of emergency
  and locks the operating frequencies as low as possible to prevent
  meltdowns.

  The daemon can only reduce the power draw. It does not restore the
  frequencies nor the power limits. It should not do that either.

  Overload signals hardware failure, hopefully related to the connection,
  either the cable itself or the connectors, which is not going to
  miraculously disappear without manual intervention.

  The imposed restrictions can be reset with the nvidia-smi utility.

    nvidia-smi --power-limit 600
    nvidia-smi --reset-memory-clocks
    nvidia-smi --reset-gpu-clocks

System integration

  The install/susd.service systemd unit file provides system integration for
  the bin/susd daemon. It expects the binary at /usr/local/bin/susd location.

Testing the functionality

  The gpu-burn utility (https://github.com/wilicc/gpu-burn) was used to
  simulate extreme computational loads.

Considerations

  The ATX 3.1 specification requires that the 12V-2x6 wires to handle currents
  in the neighbourhood of 9.2 (A).

  The RTX 5090 cards are rated for a maximal power draw of 600 W. With balanced
  distribution of load across the wires this amounts to 8.5 (A) per wire.

  Faulty connections and wires can move the load distribution out of balance,
  overloading some of the wires. To counter these in live systems it is
  possible to limit the power draw of the card.

  • The power management facilities of tested RTX 5090 cards permit dynamic
    setting between 400 and 600 (W). Within this range the card adjusts its
    clocks to meet the power limits.
  • Outside of this range the power can be further limited by adjusting the
    clocks manually, making it possible to further reduce the power draw down
    to approximately 60 (W).

References

  • Sensor interaction (smbus)
	  via https://github.com/Timic3/astral-power-monitoring
  • Sensor address and register
	  via https://long-cat.net/gitea/moosecrap/evga-icx
	  via https://github.com/LibreHardwareMonitor/LibreHardwareMonitor

Contributors

  • https://github.com/Ching367436
  • https://github.com/HotKeyIt
  • https://github.com/unknown321