SPN-based hashing

I was thinking, should a SPN with byte sub, add round key, and
complete mixing replace the unbalanced Feistel network in an
MD4/MD5/SHA-like hashing algorithm, would the basic hashing algorithm
as a whole provide a better performance that the unbalanced Feistel
version given enough rounds are used to make the security equal to its
unbalanced Feistel counterpart?

The padding steps and appending of length are retained, and MD buffer
initialisation is kept, however the processing (step 4) is replaced
with an SPN with

Byte Sub: replace each byte in an 8x8 S-box (possibly key-dependent?)
Add round key: XOR or add in the message bytes (assuming padding and
append length has been done)
Complete mixing: a high-diffusion step in which every input bit in
each of the 4 or 5 32-bit words can affect every output bit
(presumably through an MDS to mix the five 32-bit words and another
MDS to mix the bytes in each word)

So, given enough rounds to make such a scheme (say a 128-bit hash) at
least as secure as MD5, what would its performance be compared to MD5?

--
Benjamin Choi