Sequential Depth Comparison

For a length-512 sequence, how many sequential processing steps must a vanilla RNN execute, and how many sequential token-wise steps does a standard full-sequence Transformer need at inference once the whole block is available?