Søren Sandmann <sandmann at cs.au.dk> writes: > Intermediate results from matrix and vector multiplications fit in 64 > bits, and That is, as long as the final result doesn't overflow the 31.16 type, the intermediate values fit in 64 bits. Søren