Phoronix reported an approximately 2% aggregate performance difference between 8,800 and 12,800 MT/s settings in a retest of an AMD EPYC 9686F 2P system, published November 16, 2026. Individual workloads varied: some had much larger differences, while several changed little. (Phoronix’s benchmark results)
For server planning, use the aggregate as a broad reference and check results for the workload you actually run.
The test changed memory speed, not the modules
The system used two AMD EPYC 9686F processors and 32 SK Hynix HMCG98CHJMAA0EN MRDIMM-12800 modules, each with 64 GB of capacity. Phoronix compared target speeds of 8,800 and 12,800 MT/s—million transfers per second—by changing the firmware setting on the same modules. (Phoronix’s test configuration)
Workload made the difference
Across dozens of benchmark results, the reported overall difference was about 2%. The individual results varied considerably: Phoronix reported differences of 26% in HPCG and 20% in BRL-CAD, while the results for LIBXSMM and OpenFOAM varied with the test configuration. (Phoronix’s workload results)
| Workload or result | Reported difference between settings | Condition |
| Overall benchmark results | About 2% | Composite across dozens of results |
| HPCG | 26% | — |
| LIBXSMM | 11–25% | Depends on matrix size |
| BRL-CAD | 20% | — |
| OpenFOAM | 5–17% | Depends on mesh size |
| WRF | 11% | Weather-forecasting workload |
Phoronix described little difference in OpenVINO and OpenSSL; most Blender scenes and compilation tests also changed little. (OpenVINO and LIBXSMM results · OpenSSL and Blender results · Compilation results)
The comparison is between two firmware speed targets on the same MRDIMM-12800 modules, rather than separate MRDIMM-8800 modules or DDR5 RDIMMs. When lower target speeds were selected, the system reverted to 12,800 MT/s after boot.