LAPACK/ScaLAPACK Development

When the scale of the matrices are small, I found that reference blas is much faster than those optimized blas. By small scale I mean the size of the matrices are small than about 30x30. I tested several Libs: reference blas, GotoBlas, ACML and Atlas, and found that reference blas on netlib is the fastest. Can anybody here tell me why? Thank you

neodreamer wrote:When the scale of the matrices are small, I found that reference blas is much faster than those optimized blas. By small scale I mean the size of the matrices are small than about 30x30. I tested several Libs: reference blas, GotoBlas, ACML and Atlas, and found that reference blas on netlib is the fastest. Can anybody here tell me why? Thank you

Neodreamer,
hmmm, this doesn't make much sense to me. However, measuring the performance of an operation on such small input data is quite a tricky thing; you have to be aware of timer resolution and cache effects, Can you tell me more about the way you measure it?

alfredo