> All of the BLAKE variants on ARM significantly underperformed Intel (between 20--80% of Intel’s performance) and we believe that this is because the underlying library being used has no ARM-specific optimizations but it does have optimizations for Intel platforms with AVX2/SSE4.1.

What would ARM-specific optimization look like, in this case? I'm not intimately familiar with the Graviton architecture, but I'm guessing their best option would be NEON-based vectorization which can be quite slow relative to AVX (last I checked).