Cloudflare's Gen 13 servers: trading cache for cores for 2x performance

14 hours ago

Cloudflare deployed its 12th Generation server fleet two years ago, based on AMD EPYC™ Genoa-X processors with 3D V-Cache.
The transition to FL2, a Rust-based rewrite of Cloudflare's core request handling layer, allowed unlocking the full potential of Gen 13 hardware.
AMD EPYC™ 5th Gen Turin processors offer 2x core count, improved IPC, better power efficiency, and DDR5-6400 support but with reduced per-core cache.
FL1, the legacy software stack, faced significant latency issues due to reduced L3 cache in Turin processors.
FL2's architecture, with better memory access patterns, reduced dependency on large L3 caches, enabling linear performance scaling with core count.
FL2 on Gen 13 achieved 2x throughput, 70% lower latency, and 50% better performance per watt compared to Gen 12.
Gen 13 servers with AMD Turin 9965 processors are now being deployed globally, offering up to 100% higher throughput and improved power efficiency.

Hasty Briefsbeta