Intel Xeon 6 Granite Rapids Memory Scaling Performance From 6 To 12 MRDIMMs
NAMD with a million Atom model topped out on the Xeon 6980P with eight MRDIMM-8800 modules.
The QuantLib quantitative finance software continued scaling up to 12 MRDIMMs but less effective with eight MRDIMMs being rather suitable for achieving most of the performance at a lower cost with today's memory pricing.
The AMG multi-grid solver is one that always loves memory bandwidth.
Interestingly with the libxsmm library was a performance drop at eight and ten MRDIMMs but returned to scaling great at 12 MRDIMMs for this library for specialized dense and sparse matrix operations and deep learning primitives.
The Laghos benchmark configurations tested stopped scaling after eight MRDIMMs on this Granite Rapids server.
The HeFFTe library scaling performance also dropped off after eight MRDIMMs for most configurations.
Similarly for the NWChem simulator with the C240 input wasn't any further gains beyond eight MRDIMM-8800 modules.