How Cloudflare Got 2x Faster Servers With Less Cache
#ThisWeekinNET — Episode 134 In this episode of This Week in NET, JQ Lau and Victor Hwang from our Network & Infrastructure Strategy team walk us through Cloudflare's 13th generation of servers — the machines that power a significant part of the internet across 330+ cities worldwide. The Gen 13 program doubled compute density by jumping from 96 to 192 cores, but that came with an 83% drop in L3 cache. The team explains how a bold hardware bet, combined with Cloudflare's FL2 Rust-based software rewrite, turned that trade-off into a win across throughput, latency, and power efficiency. From counterintuitive fan physics to credit card pen tests on chassis intrusion switches, this conversation covers the full stack: CPUs, memory, storage, networking, security, and what's next — including post-quantum readiness at the hardware layer. Check the Cloudflare Blog: https://blog.cloudflare.com/gen13-launch/ https://blog.cloudflare.com/gen13-config/ 🎧 Subscribe for weekly conversations about the Internet and Cloudflare: https://ThisWeekinNET.com ⸻ Timestamps 00:53 — Blog recap: what Cloudflare announced (including agents can now actually create Cloudflare accounts, buy domains, and deploy) 02:40 — Meet JQ Lau and Victor Hwang 03:52 — From Gen 11 to Gen 13: the evolution of Cloudflare's servers 05:04 — Doubling compute power while cutting cache by 83% 06:54 — The journey to choosing the right CPU 10:04 — Scratchpad vs bookshelf: cache and memory explained 12:08 — Why 192 cores won over 128 cores 15:35 — FL2: Cloudflare's Rust-based software rewrite 18:12 — Hardware and software co-design: why neither works alone 18:37 — Memory, storage, and networking upgrades 22:18 — Dual GPU support and future accelerators 23:25 — Inside the Gen 13 chassis: what changed visually 24:51 — Why adding a 5th fan saves power (counterintuitive physics) 25:59 — Server security: memory encryption, PCIe encryption, intrusion detection 30:12 — 50% better performance per watt and what that means at scale 33:54 — The Austin lab: where hardware gets tested before production 35:10 — How AI helped design Gen 13 37:13 — 500 terabits per second: Cloudflare's network milestone 38:30 — What's next: Gen 14, rack-scale design, and post-quantum hardware 41:16 — Supply chain planning: lessons from COVID and the AI buildout