There is only one error which goes by undetected by the mining software and it's a not-found valid nonce. But the end effect of such error is nothing more than slightly lost hash rate, i.e the same as hashing completely wrong and returning invalid nonces.
So yes bitcoin mining is very different from CPUs when it comes to error tolerances. The silicon is designed with performance as priority and fault tolerance is spent on... perforance by multiple redundancy. If you look at a bitcoin mining asic then it has a very large number of hashing cores, each relatively simple. It is expected that likely not all parts of all chips will function optimal, but that is outweighted by the parts that do work, all together delivering bitcoin hashing performance.
Being a competitor I won't comment on the validity of Hashfast performance numbers other than that their numbers do not match with our thermal calculations. Will be very interesting to see what is reported in the field when delivered to customers outside the lab bench.