PinnedInGoogle Cloud - CommunitybyFederico Iezzi·Mar 261 Million Tokens Per Second: Qwen 3.5 27B on GKE with B200 GPUsFrom 22K tok/s on 4x H100 to 1M+ on 96 B200s. Every failure included.A response icon1A response icon1
PinnedInGoogle Cloud - CommunitybyFederico Iezzi·Jul 26, 2025Scaling Inference To Billions of Users And AgentsDiscover how Google Cloud scales LLM inference to billions of users with GKE, TPUs, vLLM, and advanced networking for next-gen AI AppsA response icon1A response icon1
PinnedInGoogle Cloud - CommunitybyFederico Iezzi·Nov 16, 2024Google Axion is a game-changer — let me show you whyIn April of this year, at Next ’24, Google introduced Axion, its first custom ARM-based CPU built specifically for datacenter workloads…A response icon2A response icon2
PinnedInGoogle Cloud - CommunitybyFederico Iezzi·Apr 21, 2024Forwarding over 100 Mpps with FD.io VPP on x86Behind the curtain of every website load and social media feed lies a vast, unseen network of Ethernet packets. These tiny data carriers…A response icon3A response icon3
InGoogle Cloud - CommunitybyFederico Iezzi·Oct 22Stop clicking “Yes”: shipping an agent permission policy inside the repository with Google…· 1. The problem · 2. Why “just ask the user” fails · 3. What Antigravity supports today · 4. The design ∘ .agents/hooks.json ∘…A response icon4A response icon4
InGoogle Cloud - CommunitybyFederico Iezzi·Sep 10The Idle Half of a GPU PodWhat happens when you actually use the 384 vCPUs sitting next to eight Blackwell GPUs? Buy a GPU node and you are buying a CPU node too…A response icon1A response icon1
InGoogle Cloud - CommunitybyFederico Iezzi·Sep 2Disaggregated Serving on B200s: 5.9x goodput, and what it costsPart 1 of this series ended with a confession. Every disaggregated number in it was measured on hardware that could not do the job…
InGoogle Cloud - CommunitybyFederico Iezzi·Aug 18Your KV Cache Transfer Is 100x Slower Than It Should BePrefill/decode disaggregation on GKE with RTX PRO 6000, and the one config value nobody documents.A response icon3A response icon3
InGoogle Cloud - CommunitybyFederico Iezzi·Aug 12One core, nearly 6 Mpps: testing AF_XDP on Google Cloud C4NAt Next ’26 this year, we announced our first network-focused machine type 🎉. Follows a quote from the official announcement:
InGoogle Cloud - CommunitybyFederico Iezzi·Jun 12Enabling Claude Fable 5 on Agent PlatformWhat’s happening