Why does the search limit NB to 80?
The default ATLAS search limits GEMM’s blocking factor to at most 80. On systems where larger NB actually blocks for the L2, blocking for the L2 prevents ATLAS from using it’s multilevel blocking parameter, CacheEdge. In this case, larger blockings may result in superior kernel timings (which do no L2 blocking), but if an L1-contained NB is used, similar or superior performance may be obtained in full GEMM with a tuned CacheEdge. In this case, the GEMM speedup is illusory, but the application and small-case gemm slowdown (discussed below) is quite real. On machines with large L1, or very fast L2, GEMM may indeed get a asymptotic speedup from larger blocking factors, but it is still almost always a bad idea, as outlined below.
