[...] I can't find any performance numbers. I'm inclined to believe this is more of a 'secure box' than a performance contender.
You would be hard pressed to try and scale this to a point where it was interesting, imho.
I skimmed through the datasheet for the AT88SA102S ... when I was tired...
It does report tEXEC_MAC
= 30ms (Delay to execute MAC command) in table 4-1. I think this is the only SHA-256 operation where you get the output, the other is to generate some encryption key for burning in the secret.
It looks like the 512 bits it hashes are rather tightly constrained - you can include some message, but the rest is internal (secret, write-only) state. But assume it will hash anything for you, on request. Input is 36 bytes, output is 32 bytes; each(?) plus 3 bytes to form the IO Block.
Sending data down the serial interface (tBIT
in table 3-1) takes >37us per bit. Receiving data takes >41us (the chip's clock spec is looser than that of the host issuing commands).
Total time per hash (MAC command) looks like
IO: (36 + 3) byte * (37 microsec / bit) + (32 + 3) byte * (41 microsec / bit) = 23ms
+ Compute: 30ms
Per chip looks just under 19 hashes/sec. You need 150,000 of these to match a modern (3Mhash/sec) CPU.