In 5 years all ASICs will be replaced by this...

[DeathAndTaxes]

The question isn't necessarily how fast it can compute anything, it's how fast it can compute SHA256. By the time this is out, there will be better ASICs that will be faster for mining than this thing.

Exactly.  This is rather unremarkable compared to the power of existing ASICs.  The article doesn't mention integer performance and that isn't unusual most super computers are focusing on high floating point performance and high bandwidth interconnects. 

Still if we assume an equivalent integer performance (more likely it will be significantly lower) that would be 10 trillion IntOps per second.  SHA-256 is 80 rounds lets assume (highly optimistically) that it can do one round per IntOp that would be 80 IntOps per hash.  Lets round up to 100 to include loading, incrementing, checking logic.   So 10 trillion integer operations = 100 GH/s.

[Come-from-Beyond]

The question isn't necessarily how fast it can compute anything, it's how fast it can compute SHA256. By the time this is out, there will be better ASICs that will be faster for mining than this thing.

Exactly.  This is rather unremarkable compared to the power of existing ASICs.  The article doesn't mention integer performance and that isn't unusual most super computers are focusing on high floating point performance and high bandwidth interconnects.  

Still if we assume an equivalent integer performance (more likely it will be significantly lower) that would be 10 trillion IntOps per second.  SHA-256 is 80 rounds lets assume (highly optimistically) that it can do one round per IntOp that would be 80 IntOps per hash.  Lets round up to 100 to include loading, incrementing, checking logic.   So 10 trillion integer operations = 100 GH/s.

Difference between floating point and integer is not so big. Look at SSE registers that work with the both representations using the same bits but slightly different opcodes.

Edit: Even more, 1 floating point operation equals to 4 integer operations.