The Performance Impact Of Genoa-X's 3D V-Cache With The AMD EPYC 9684X
One of the real-world workloads with the most significant performance impact from AMD 3D V-Cache with Genoa-X is, of course, the OpenFOAM computational fluid dynamics software. The shear difference from toggling 3D V-Cache for OpenFOAM was very profound and speaks to the possibilities of 3D V-Cache. Other CFD software can also benefit too from 3D V-Cache while of course my benchmarking is focused on the open-source workloads.
Having 3D V-Cache enabled (default) led to a CPU power consumption increase by around 10 Watts on average, or about 3% more which isn't bad at all considering the significant time savings.
With the large mesh size tested with OpenFOAM there can be significant time savings from the 3D V-Cache.
The libxsmm library for specialized dense and sparse matrix operations and deep learning primitives saw huge uplift too from AMD 3D V-Cache, similar to the nice improvements observed too during the Intel Xeon Max testing with HBM2e.
The AMD EPYC 9684X CPU power consumption was also slightly lower when leveraging the 3D V-Cache for this benchmark.
For other data-set sizes tested, libxsmm continued to show nice benefits to the 3D V-Cache.
The HeFFTe library also showed the very significant difference from toggling 3D V-Cache.