BQ78350: Why does the remaining capacity drops suddenly from 75% to 7%?

Hi Feiyue,
I'm not sure I understand, but please see if this helps.
If you have a 3.7V nominal cells with 2300 mAh capacity:
The measured capacity will typically be a little more when new and less when old, and it will depend on the discharge rate.
A single cell power capacity would be 2300 mAh x 3.7 V = 5290 mWh capacity.
When you stack 5 cells in series to make a 5S1P pack, you still have 2300 mAh capacity since the current flows through all cells. But you now have 2300 x 3.7 x 5 = 26450 mWh capacity.
Your design capacity may normally match the connected cells, or might be set differently for some margin.
For 11000 mAh design capacity with the 2300 mAh cells, you might parallel 5 cells. This would understate the 5 x 2300 = 11500 mAh expected capacity. The pack would be commonly referred to as a 5S5P pack. Learned capacity can be initialized to the design capacity but will be learned after a qualified cycle.

The bq78350 determines capacity in 3 ways:
First at reset it will estimate the capacity based on the voltage. The proper chemistry must be loaded for that to be meaningful so that the voltage matches the % full for that cell type.
Second if charged to charge termination and configured to do so the gauge will set RSOC to 100%. It is fully charged.
Third if discharged to the EDV2 point it will learn capacity and update the learned capacity. There are learn up and down limits which will restrict the amount of capacity change per cycle. The RSOC will be set to 7% (default) when it learns. The EDV2 point will be the fixed voltage if configured that way, or calculated by the gauge if using CEDV. If using CEDV the parameters entered should be calculated from the GPCCEDV tool. If using the default CEDV parameters the learning may not represent the real cell performance.

Reading your post again I expect you are testing with a 5S1P pack, in this case the design capacity (and initial learned capacity) should be 2300 mAh. RSOC would count faster as the pack is charged and discharged, so there should not be as significant of a jump at full charge or when learning. Optimizing the parameters should minimize any jumps. The 1.6A load seems consistent with a 1P pack, but if you do need higher capacity for longer run time you will need to add parallel cells.
If this does not help please reply or post again.

Hi Feiyue,
The cWh is used, centi-Watt instead of milli-Watt. Yes it is divided by 10. It allows better use of the data range since Wh are multiplied by voltage and will be a larger number.
Yes, 4810 seems correct.