It's not going to be tomorrow.  SA5 has no more easy optimizations left; the current version maxes out the memory bandwidth of the card.  Additional significant performance improvements are going to require big changes to the algorithm.  My top idea involves reducing NR_SLOTS to 8 (or maybe even 6 or 4) from 12, and adding a collision overflow table for each round.  By reducing the table size it reduces the time to modify the table (shifting the indexes and updating the XOR'd hashes) and could improve performance by 50-75%, putting SA in the performance range of OptiminerZcash.  This idea would require changes in almost every part of the kernel including ht_store, equihash_round, and expand_refs, and would require many hours of coding.