Yes, NUMA and large numbers of core will pose completely new challenges for software optimization. Applications will have to become aware of the memory architec

Yes, NUMA and large numbers of core will pose completely new challenges for software optimization. Applications will have to become aware of the memory architecture of the underlying machine. They will have to make explicit assignments of memory allocations and threads to NUMA nodes based on their domain specific needs. In some cases, even duplicating data structures may be the right call. This is going to challenge how most developers write fast software. The other thing is that even seemingly trivial things like spawning and synchronizing with lots of threads will be much more complex on CPUs with many cores. At some point, naively looping over all threads is going to be too slow. I think that the limit is going to be around 64 cores. Past that, you should actually parallelize your worker thread management to stay efficient. There is precendent for this in HPC, e.g. MPI implementations.