The author conducted tests to analyze latency under load, similar to processes described in Ampere’s Hot Chips 2024 presentation. Various AMD chips are sensitive to thread placement, affecting latency. The system uses multiple interconnects for high core counts, with each Zen core sharing an L3 cache within a cluster. AMD’s Infinity Fabric provides flexible interconnectivity. Zen 4 and Zen 5 exhibit improved behavior under high loads compared to Zen 2. AMD potentially enhanced traffic management policies to enhance performance in Zen 5. Results suggest Zen 5 outperforms Zen 4 in latency management under load due to faster memory and Infinity Fabric.
https://chipsandcheese.com/p/pushing-amds-infinity-fabric-to-its