You can make cache coherence scale to multiple cores with precise tracking and the overhead of precise tracking can be reduced by arranging your cores as clusters.

Exact tracking basically requires a bit field with one bit for each core. The problem is with the storage overhead since a 64 core cpu would need 64 bits per cache line. However if you had a 64 core CPU you could partition it into 8 clusters with 8 cores each which means you only need 8 bits per cache line for exact tracking.

http://www.cis.upenn.edu/acg/papers/cacm12_why_coherence.pdf