Mem usage

[Tesla32X]

First of all, thanks for great update. It looks very promising!

I did some testing after you released this yesterday, and wonder if this is normal behaviour:

Quick96 is faster on 1 GPU than on 2. (Actually it seems twice as fast in single mode, around 100 ms/it on Gtx 1080ti, but then 220 ms/it when I use both cards)

When running SAEHD with two cards (identical), only memory from 1 card is used. And on average it seems to be using only half of what is available looking on the graph. When upping batch size I still get OOM.

Here are GPU loads, mem usage etc.

[iperov]

is sli disabled?

[iperov]

I already tested new dfl on mining farm, all works fine.

[iperov]

Quick96 is faster on 1 GPU than on 2

it should not be faster. MultiGPU just divides batch across gpus.

[Tesla32X]

Ah. I restarted and it all seems fine now. My bad.