Question regarding NUMA scaling and UPI bottlenecks for heavy FLIP/Pyro sims

[nenegad149]

Hi everyone,

I’m currently planning a refresh for my home simulation node (mostly for heavy FLIP and high-res Pyro work in H19.5/H20) and looking into the secondary enterprise market for a dual-socket build.

I know Houdini loves single-threaded performance for the viewport and SOPs, but for the heavy lifting, I need parallel throughput. I’ve been looking at a specific configuration listed as CPU & Processors /Xeon 16 Core/2.9GHz-10.4GT-UPI.

My main concern is the interconnect speed. In a dual-socket environment, has anyone experienced significant slowdowns due to NUMA context switching with the 10.4GT/s UPI link? I’m trying to figure out if the 2.9GHz base clock across 32 total threads (assuming a dual setup) will be bottlenecked by the interconnect when pushing massive attribute arrays between sockets, or if Houdini’s memory management is efficient enough now to keep the cores fed without saturating the QPI/UPI bus.

Any real-world experiences with similar Xeon architectures would be appreciated.

Thanks!