42 ms·
This uses less memory so you can do fine tuning or hardware with less vram but at a cost of taking longer on training - there is a throughput penalty, the paper
by Scipio_Afri 2y ago
This uses less memory so you can do fine tuning or hardware with less vram but at a cost of taking longer on training - there is a throughput penalty, the paper detailing the technique shows something like a 15% decrease in throughput.