Today I had a thought. Bulletproofs are good at batching, so why not try batching the amounts themselves? E.g. split the 64 amount bits into two 32 bit numbers, and batch those. Conclusion after perf testing: it's slower.

The reason I thought it might be faster is because you can reduce the range that needs to be proven.

<emsczkp:matrix.org> I did a research in the past that shows that proving 1 rp with batched amounts is better than multiple rp for each amount, at the same bit domain

emsczkp: Unfortunately there’s no opportunity to do more of that. Tx amounts are already internally batched.