Best method is to first run lc0 backendbench to set best minibatchsize
Then run lc0 benchmark 34 positions and determine nps>
Example
```.\lc0.exe benchmark --threads= --minibatch-size=136 --nncache=20000000 --max-collision-events=917 --backend=multiplexing --backend-opts="(backend=cuda-fp16,gpu=0)” --max-prefetch=32 --max-collision-visits=1000 --max-out-of-order-evals-factor=2.4 --multi-gather=true --num-positions=34```
Then run SF bench
Example
```bench 16384 36 10000 current movetime```
Try different ram and number of threads depending on your setup
Determine SF NPS
Ideally SF NPS / Lc0 NPS should be in the ratio 1 : 700 - 1000 (With a T60 net)
You can achieve this ratio by increasing / decreasing the number of threads for SF
Lothar