MRDIMM 8800MT/s vs. DDR5-6400 Memory Performance With Intel Xeon 6

For memory intensive HPC applications, MRDIMMs were the real deal. Very impressive uplift with MRDIMM 8800 over DDR5-6400.

One of the only workloads with an apparent regression was the BRL-CAD software. With this CAD app, DDR5-6400 memory was faster than the MRDIMMs -- perhaps due to tighter memory latency.

The MRDIMMs were helpful for the OpenFOAM CFD software. With DDR5-6400 the Granite Rapids mesh time performance was similar to EPYC Genoa(X) but with MRDIMMs allowed it to come out in front.

Particularly for memory bandwidth intensive HPC workloads and other technical computing workloads, the MRDIMM 8800 modules proved to deliver very nice performance benefits over DDR5-6400 memory. With HPCG there was a 33% improvement simply going for MRDIMMs rather than DDR5-6400. The AMG benchmark was nearly 30% faster. OpenFOAM CFD was up to 15% faster... Very nice gains for Incompact3D, WRF, DSP, and other workloads. I was very impressed by the MRDIMM 8800 performance for HPC workloads.

Micron is said to be bringing to market MRDIMMs from 32GB to 256GB per module. It will be interesting though to see what MRDIMMs end up costing in the market place in both the near term and a bit further out. For HPC workloads at least the MRDIMMs will likely prove to be of great value. My Intel Xeon 6 Granite Rapids benchmarking is continuing, so stay tuned as I continue exploring more areas of the Xeon 6980P and investigating other workloads that may benefit from MRDIMMs. Thanks to Intel for supplying the review hardware and all the Micron 64GB 2RX4 PC5-8800X-HA0-1110-XT / Micron 64GB 2RX4 PC5-6400B-RA1-1211-XT DIMMs to make this testing and performance benchmarking possible at Phoronix.