Have you checked if SNAP is able to use all of your 128 GB? It is described here: Gpt performance and here: S1 data calibration error
I would furthermore suggest to split the graph into two parts. Graphs which are too long have problems passing the intermediate products to the next operator. That releases used memory and is more effective in terms of computing time: Coregistrating more than two Sentinel-1 SLC IW products and Problems obtaining an interferogram from two product sets