0.010.1110100101001000peak (vector + FMA)no FMAscalar (no vectorisation)memory boundslope 1 = 204 GB/scompute boundstream triad92% of roofsparse matvec91% of roofstencil, 7-point90% of roofFFT, 1M point75% of roofdense GEMM (naive)6% of roofdense GEMM (tuned)83% of roofN-body90% of roofarithmetic intensity (FLOP per byte of DRAM traffic)attained performance (GFLOP/s)Roofline: log-log is what makes both limits straight and the ridge a cornerpeak (vector + FMA)no FMAscalar (no vectorisation)