HiveOS Update:
wget https://github.com/Lolliedieb/lolMiner-releases/releases/download/1.95a/lolMiner_v1.95a_Lin64.tar.gz && tar -xvf lolMiner_v1.95a_Lin64.tar.gz && pidof -x SCREEN && miner stop || echo lolminer && cp -adpR 1.95a/lolMiner $(ls -d1 /hive/miners/lolminer/*/ | tail -1) && miner start
wget https://github.com/Lolliedieb/lolMiner-releases/releases/download/1.95a/lolMiner_v1.95a_Lin64.tar.gz && tar -xvf lolMiner_v1.95a_Lin64.tar.gz && pidof -x SCREEN && miner stop || echo lolminer && cp -adpR 1.95a/lolMiner $(ls -d1 /hive/miners/lolminer/*/ | tail -1) && miner start
Edit: Windows version re-released to allow negative --keepfree values.
The continuous mining feature requires to use as much memory on the GPU as possible. Having a larger fraction of GPU memory available will allow to put more of the DAG in this space and therefore increase performance. That said: trying to allocate too much GPU memory will result in memory errors and the miner failing to start up or start the build.
The internal detection of the miner is tailored for GPUs running in a Linux system with no screen connected and driver version 535. Therefore this value might be good for your purposes, but you might want to tweak it. In order to do so there is the parameter --keepfree that is an offset to this detection in MBytes. If your GPUs do not start up, then try to increase the value of keepfree until they do. We recommend steps of 8 MBytes in this case. Note that we found out that driver version 525 and 535 seem to allow a larger memory used then newer ones like 550. Therefore we advice to use Cuda driver 535.x.
In particular on Windows systems the automatic detection might underestimate the available memory by a lot, especially if a screen is connected and parts of the GPU memory are used for display purposes. While during the startup of the miner the driver may report rather low proportions of GPU memory being free it can happen that the actual possible space available to the miner is larger, because the driver may move less frequently used memory patches to the system memory. In this case you might want to use a negative value for the --keepfree parameter in order to allow the miner to use more memory then reported free by the system. On an internal test with a 3070 running Windows 10, this value can be as high as -512. Other cards tested were better with a value of -400. We recommend starting near 0 and then modify up / down this value in steps of 16 (initially) and 8 (later) to tweak for optimal performance. Note that increasing the value will improve stability while reducing it will improve the speed.
Note that keepfree works in two ways. If you specify only a single value, this value will be applied to all GPUs in the system. You can give multiple values to all GPUs in the system by providing a comma separated list of values.
