It seems that the article goes a long way to avoid telling us how the Tidescale technology intends to solve the inherent problems when building virtual NUMA clu

It seems that the article goes a long way to avoid telling us how the Tidescale technology intends to solve the inherent problems when building virtual NUMA clusters with commodity hardware - latency and throughput (and, as a consequence, memory/cache coherence) when accessing memory that is located in another machine. The whitepaper linked in the article (https://www.tidalscale.com/wp-content/uploads/2020/03/Wandering-CPU.pdf https://www.tidalscale.com/wp-content/uploads/2020/03/Wander...) is also not really detailed. It uses the "wandering detective" analogy, which - I think - implies the idea of moving the code to where the data is located. However, this probably doesn't help when the computing capacity required for processing that data exceeds the capacity of a single machine. The principle of virtualizing NUMA access over different machines worked fairly well using specialized hardware, e.g. using SGI's NUMAlink interconnect on MIPS machines, but it's not obvious from the article that special hardware is required for the Tidescale approach.