It's a little bit of a misrepresentation on the part of the presenter. BLAS is a specification for linear algebra software, of which implementers are free to us

It's a little bit of a misrepresentation on the part of the presenter. BLAS is a specification for linear algebra software, of which implementers are free to use the most appropriate representation of numbers as they wish - the fastest being floating point because of native hardware support. Although the naming schemes for BLAS routines do explicitly reference "single precision" and "double precision" so now I'm confused too.