Show Notes
GPU inference throughput depends on more than accelerator generation or count. Memory bandwidth, model parallelism, cache configuration, and the load generator itself all influence measured throughput. Federico Iezzi, Customer Engineer at Google Cloud, explains how his team achieved 1 million output tokens per second using Qwen 3.5 27B, vLLM, GKE Autopilot, and NVIDIA B200 GPUs. The discussion covers: Why memory bandwidth limits decode performance
How Federico chose between tensor and data parallelism
What changed after enabling multi-token prediction and reducing the KV cache footprint with FP8 quantization.
Sponsor This episode is sponsored by LearnKube. Download the free book, The Technical Guide to Kubernetes Rightsizing, to understand what Prometheus and Grafana cannot tell you about safely reducing requests and limits. More info Find all the links and info for this episode here: https://ku.bz/1xD9Md0mb
Interested in sponsoring an episode? Learn more.
More from KubeFM
Non-deterministic agents pose specific challenges for platform teams in observability, state management, governance, and trust. Mauricio (Salaboy) Salatino explains why agentic applications behave like distributed multi-agent systems. His test assigned agents to take an order, cook the pizza, deliver it, and charge the customer. One order crossed 15 containers and produced 200 traces. In this interview: Why agent frameworks can recreate monolith scaling and resource contention
KubeSelect is a new show that tests one bold Kubernetes hypothesis with one expert. Hosts Salman Iqbal and Bart Farrell examine the evidence and ask whether the claim holds up. In episode two, Kat Cosgrove tackles a long-running question: should teams run databases on Kubernetes? StatefulSets, persistent volumes, CSI, and database operators changed the technical answer, but they did not remove the operational risk. In this episode: How Kubernetes storage has matured and why old assumptions persist
Kube Select takes one bold Kubernetes hypothesis and tests it with an expert. In the first episode, Salman Iqbal and Bart Farrell ask whether most organizations have an observability problem or a decision-making problem. Henrik Rexed, Senior Staff Engineer at Dynatrace, challenges the original claim. Teams can collect metrics, logs, and traces, but that telemetry needs system relationships, ownership, and deployment history to support a confident decision. In this episode: Why telemetry without context slows incident response
Kube Signals starts where the keynote ends: with the trends that platform teams will have to operationalize next. In this special episode, Brian Teller speaks with Saiyam Pathak about his KubeCon India keynote and the shift from developer platforms to AI factories. They examine what GPU scarcity, shared accelerators, and AI workloads mean after the conference slides meet real infrastructure. In this interview: Why GPU infrastructure is becoming a platform-engineering concern
At enterprise scale, a deployment pipeline that runs Helm upgrades directly against Kubernetes hides drift, mixes configuration with CI logic, and makes the last pipeline run the source of truth. Elad Cohen explains how WSC Sports moved from Azure DevOps to GitHub Actions and redesigned delivery around Git and Argo CD. The resulting platform separates builds from deployments, keeps service configuration in values files, and continuously reconciles clusters. In this interview: Why CI should change Git instead of the cluster