No real throughput data to choose RAM capacity vs. RAM bandwidth for MoE offload — expensive hardware bets made blind

Summary

Users who want to run MoE models larger than VRAM (with routed experts offloaded to system RAM) must decide between more RAM capacity and faster RAM — e.g. 96GB DDR5-6000 vs 128GB DDR5-5600, or staying on 8-channel DDR4-2933 vs rebuilding on DDR5-6400 Threadripper. This is a multi-thousand-euro, platform-level decision, yet there is no trustworthy source of real decode/prefill tokens-per-second for offload across memory topologies, no model of how bandwidth, capacity, and expert-caching interact, and no way to know whether ~2× bandwidth is worth ~3× price. Users are left asking strangers on Reddit for anecdotal t/s numbers and risking money and months of rebuild time on the wrong platform. Existing model hubs and benchmark suites report GPU-only or aggregate scores and never map a specific model + quantization + offload configuration to expected throughput on a given memory subsystem — so the hardware-planning side of the quantization decision is completely unsolved.