To make a long story short, the AES3 interface limits word length to 24 bits. This means the noise contribution of the wordlength is at -146dB. AD/DA converter chips with SNR hovering around 125dB can be had these days from companies like ESS and ARDA. This means that the 24 bits word length is not yet even close to being an important contributor.
Noise performance is limited by a variety of factors, most are purely analogue in nature (thermal noise, settling time errors, matching errors, jitter). It's a matter of economics that an optimum is sought between all noise sources, so there's never a situation where one noise source dominates or another one is wholly negligible. The only exception is quantisation noise where noise shapers are easily designed to contribute far less noise than anything else in the circuit and internal word lengths are chosen 4 or more bits longer than necessary.
This means there isn't a platonic "perfect conversion" sitting somewhere in an otherwise noisy circuit. Everything adds its own flavour of error.