So, I should’ve posted here first, but the data is here

shared mem is significantly smoother, and the bottleneck shifted to reading the model data. Let me make that async, and try again. (Also, there are a few bad design choices that I made that should be fixed too)