3. Some results

The results presented here where done on a dual PII 400 system running Linux The time measurements are done using the time stamp counter of the processors. The base BLAS libraries used are from the Fortran 77 implementation⁶, ASCI Red project and ATLAS project (for level 3). The memory bandwidth measurements are done with the hardware performances counter available on the Pentium Pro family processors⁷. For each BLAS we test we compare single and dual performances in two case, with cache memory flushed out (datas are read from main memory) and with cache memory loaded: the test is done several times before measuring execution time (remark: this does not seem that data will fit into cache).