nprofile1qy2hwumn8ghj7un9d3shjtnyd968gmewwp6kyqpq7awtznnk6vhc4q689zf0xds2a4uydanj8u8ljcyr8m9waamquzkskf3vpz (nprofile…3vpz) I would expect shadow passes to be entirely bandwidth-limited. You're doing a very tiny pixel shader, so it's all Z and FB writes.
So adding any texture read that misses the cache - yup, direct perf hit.