Quite a lot of submissions: https://hn.algolia.com/?query=Why%20Aren%E2%80%99t%20There%20Software-Defined%20NUMA%20Servers%20Everywhere%3F https://hn.algolia.co

Quite a lot of submissions: https://hn.algolia.com/?query=Why%20Aren%E2%80%99t%20There%20Software-Defined%20NUMA%20Servers%20Everywhere%3F https://hn.algolia.com/?query=Why%20Aren%E2%80%99t%20There%2... Personally Im more interested in going the opposite way. Many of the greatest most-scalable software systems out there are already built for scale out. Rather than build careful & costly coherency into our hardware, I'd like to see non-coherent architectures. Make systems expose CXL.mem or RDMA interfaces or just plain networking and offer gobs of medium/high radix interconnect in place of fabric. Given that we have >100 core and 16-channel ram (now 32c since ddr5 is smaller 32b channnel) chips, not evem sure we need coherency across a socket, am not convinces we shouldnt just start thinking of a socket as having effectively multiple isolated chips. Certainly would be a wild trade-off, likely implies splitting the LLC, but how much faster/simpler/power power would it be to shrink the domain of that cache, or to make it non-coherent to further off domains. I really dont have the figures to back this up. I dont have a good sense of how much power draw is fabric, coherency, or how much latency the wide reach adds. Certainly coherency makes semse for many applications & we do seem quite good at it. Still, I'd be so interested to see some in-between, some tries!