-
Notifications
You must be signed in to change notification settings - Fork 597
Nvidia LHR and Kernel Mode
- Mandatory to use 470.103.01 or 510.x drivers! They have turned out to be much more stable in unlock in testing then older drivers. Guide Nvidia Drivers Update
- The miner software needs to learn certain aspects about your cards over time to optimize for the best balanced pool hash rate. The needed calibration will get better over time, but the things learned are too complex to display them or store into parameters.
- Therefore in automatic mode the miner might start with a slightly slower hash rate at the beginning, but will get better over time. In non-automatic mode you might see more re-calibrations during the first hour of mining then normal. If the number keeps too high after that point, lower the requested % value.
- As a consequence: the longer you can keep it running the better - avoid frequent restarts of the software.
- Try to keep your settings as stable es possible. Especially: Set clocks that your GPUs can keep without running into thermal throttling. Also use locked clock core with no set power limit to maintain a stable GPU core clock for ideal results. Having GPUs freely changing their clocks and sometimes running into their power limit will invalidate a lot of data collected and decrease efficiency of the unlocker a lot. Same applies to memory clock - do not change it after the calibration did start. Use --lhrwait parameter to delay the initial calibration until you are done with your settings.
- The initial configuration takes approx 3-4 minutes from the moment the DAG was build.
- Keep away other work load from your GPU as e.g. connected screens. Also CPU mining in parallel to the new unlocker is untested and might give unexpected results.
The parameter --lhrtune takes absolute % values to fix a certain percentage of unlocking. If chosen too high, lolMiner will reduce the value to the maximum possible. The value will be tied to keep regardless of the number of locks this implies. The default is "auto" in which the miner tries to balance the number of locks and re-calibrations with the speed to achieve, hash rate might change from time to time based on unlock quality heuristics. and "off" will disable the LHR Detection
Example: --lhrtune 77.8,off,auto that will try to get in GPU0 77.8% of unlock, GPU1 will be considered as NOLHR and GPU2 will try to get max unlock
--mode controls the loaded kernel and operation modes for Nvidia GPUs. AMD cards will currently ignore this parameter completely. You can set one value for the whole rig or provide a comma separated list of values to get individual modes for each card. Amd cards can use any value at the moment.
This is the often fastest kernel when you do not apply any power limit. Gives better performance especially on RTX 3080 and 3090, but not so much on the lower tier GPUs. Use this one when power draw does not matter any ways.
This is a power saving kernel an the default. Good choice for all Non-LHR Nvidia GPUs and RTX 3060 LHR V1 on the known 470.05 beta Windows driver. Also this mode will be used in glitch mode.
For all LHR GPUs that were started with one of the LHR modes and the fan glitch gets detected, the card will leave the LHR mode to unlock the full performance of the card. In most situations the appearance of the glitch triggers a one time card crash that makes the mining OS restart the mining software. The glitch detection allows to mine in LHR mode until a glitch on one card appears. On the next restart of the software the glitches cards will then mine with LHR mode disabled while the remaining cards keep on mining in their configuration.