A visual walkthrough of why the real comparison is not swap versus DRAM, but transparent page placement versus raw capacity pain. This page turns the episode into architecture maps, page-temperature heatmaps, and workload tradeoff charts rather than a text recap.
Vistara is interesting because the paper is not just a memory card. It is the whole path: server, ASIC, Linux placement policy, workload tuning, and the decision to sometimes disable CXL entirely.
The core argument is that the right comparison is transparent page placement. Toggle between a naive spill policy and a tuned hot-page policy to see which pages stay on local DDR5.
The paper's strongest deployment claim is that direct-attached CXL helps in production. The harder scientific question is whether it still wins when capacity is held constant.
The broader bet is not merely adding one slower tier. It is whether memory stops being welded to a motherboard and becomes something operators can shape across workloads, server classes, and refresh cycles.