Hi!
I need to compute A*B many times and I know that A contains about 20% of zeros which placed randomly.
Is there any way to use this knowledge to improve perfomance?
I tried to use sparse matrix, but it significantly slower.

20% of zeros is not significant. Trying to exploit them will kill pipelining and vectorization.

Thanks for reply!
But can you, which % of zeros is significant?

In my experience , assuming that the zeros are uniformly distributed you need at least 50% of zeros to start seeing a speedup.