Comment by zozbot234

You can make eDRAM using logic processes (which are currently less bottlenecked than RAM, at least for non cutting-edge nodes) but the cost is still prohibitive compared to specialty DRAM processes (even when considering the recent increase in DRAM prices). If you were doing that, you'd want to use it for compute-in-memory instead, which basically pushes NUMA to an extreme of having lots of tiny cores each with direct access to its own local DRAM.