Investment vs. yield relationship for memories in SOC

Today's SoC designs use significant outsourced IP and manufacturing and the overall SoC yield can be bounded by the yield of the embedded memories often provided by third party sources. Who is responsible for ensuring better memory yield - the memory IP provider, the foundry, the design team, the test house, the library provider? How can yield be maximized? What are the shortcomings of the current methods and what new investments, strategies, and collaboration are required at the design stage, on the manufacturing floor, and in the field ? How do these investments relate to yield improvement ?.