We have a good guess. They claim their 3.5 GH/s chip is USB powered. So there you go. 1400 MH/watt (or better), which means if the 1 TH/s Rig is using the same chips in parallel, 714 watts.
This 130nm shared wafer research ASIC can achieve 140 MH/J performing SHA-256 on a streaming input.
http://rijndael.ece.vt.edu/sha3/index.html
Keep in mind:
a) this was designed as a testbed for SHA-3 and runs at only 50 Mhz. Not exactly ideal for SHA-256.
b) it is on a 130nm platform. 65nm would be roughly 4x the MH/J. 45nm would be roughly 8x the MH/J
c) VT own data shows the chip can easily run at much higher clock speed without increasing the gate count.
d) the design is optimized for multi-round hashing which is ill suited for Bitcoin (where single nonce header is hashed once, check and discarded or returned).
e) it was designed as research project at my alma mater Virgina Tech (hardly ultra cutting edge fabrication).
Despite all those handicaps this unoptimized (for Bitcoin hashing) ancient 130nm multi-purpose shared wafer chip built at a university research fab can achieve 140MH/J. With optimization and a modern 45nm process >2 GH/J is certainly "possible".
Once again before you misinterpert. I am not buying a BFL product (or any mining hardware). I don't really care if BFL has the greatest hasher on the planet or the best scam. I agree it is foolish to give money upfront and hope for the best. All that being said you undermine your argument when you claim the possible is impossible.
Is a 3.5 GH/s ASIC running on 3.5W possible? Most certainly.
Will BFL produce it (and on time, and on spec)? Who the frack knows.
Are you sure that's right? They listed 13.76mJ/Gbits.
Hmm, isn't it wrong? I think the paper contradicts itself because Table III lists SHA-256 doing 1.51Gbps at 5.18mW measured (first Power column, not the second). That would be 5.18/1.51 = 3.43 mJ/Gbits. Why would the table also lists 13.76 mJ/Gbits?
I emailed the authors for a clarification.