Another technique is to transpose the left matrix so each dot product is scanned in row-order and hence more cache friendly.

Another one I tried ages ago is to use a single loop counter and "de-interleave" the bits to get what would normally be 3 distinct loop variables. For this you need to modify the entry in the result matrix rather than having it write-only. It has the effect of accessing like a z-order curve but in 3 dimensions. It's a bit of overhead, but you can also unroll say 8 iterations (2x2x2) which helps make up for it. This ends up making good use of both caches and even virtual memory if things don't fit in RAM. OTOH it tends to prefer sizes that are a power of 2.