Kubernetes GPU Management Just Got a Major Upgrade
Nvidia Distinguished Engineer Kevin Klues noted that low-level systems work is invisible when done well and highly visible when it fails — a dynamic that frames current Kubernetes innovations for AI. At KubeCon + CloudNativeCon North America 2025, Klues and AWS product manager Jesse Butler discussed two emerging capabilities: dynamic resource allocation (DRA) and a new workload abstraction designed for sophisticated AI scheduling.
DRA, now generally available in Kubernetes 1.34, fixes long-standing limitations in GPU requests. Instead of simply asking for a number of GPUs, users can specify types and configurations. Modeled after persistent volumes, DRA allows any specialized hardware to be exposed through standardized interfaces, enabling vendors to deliver custom device drivers cleanly. Butler called it one of the most elegant designs in Kubernetes.
Yet complex AI workloads require more coordination. A forthcoming workload abstraction, debuting in Kubernetes 1.35, will let users define pod groups with strict scheduling and topology rules — ensuring multi-node jobs start fully or not at all. Klues emphasized that this abstraction will shape Kubernetes’ AI trajectory for the next decade and encouraged community involvement.
Learn more from The New Stack about dynamic resource allocation:
Kubernetes Primer: Dynamic Resource Allocation (DRA) for GPU Workloads
Kubernetes v1.34 Introduces Benefits but Also New Blind Spots
Join our community of newsletter subscribers to stay on top of the news and at the top of your game.
Multi-Instance GPUs, with Kevin Klues and Pradeep Venkatachalam
NVIDIA and Google have teamed up to bring the new Multi-Instance GPU feature, launched with the NVIDIA A100, to GKE. We speak to Kevin Klues from NVIDIA and Pradeep Venkatachalam from Google Cloud on how and why people use GPUs, optimising instance shapes for machine learning, and why less is often more.
Do you have something cool to share? Some questions? Let us know:
web: kubernetespodcast.com
mail: kubernetespodcast@google.com
twitter: @kubernetespod
Chatter of the week Episode 64, with Sarah D'Angelo and Patrick Flynn Catching up with Patrick in Episode 148
Winthrop, Washington
Blackdown Hills, Devon
News of the week Azure App Services now available for Azure Arc Azure Arc and App Service blog posts
Other new AKS capbilities
Virtualization Review coverage
ECS Anywhere made GA by press release
AWS App Runner
Integrating Google Cloud DNS with GKE
Istio 1.10
Terraform 1.0
Grafana 8.0 and Tempo 1.0
Argo Rollouts 1.0
Kubesphere 3.1.0
Cilium 1.10
OpenSLO spec launched at SLOConf Episode 147, with Brian Singer and Kit Merker
Envoy GA on Windows
Chaos Experimentation Framework for Envoy
El Carro operator for Oracle Database from Google Cloud
Moco operator for MySQL from Kintone
PlanetScale GA Episode 81, with Jiten Vaidya and Sugu Sougoumarane
FoundationDB paper from ACM SIG MOD
DockerCon announcements
Coverage of Development Environments from The Register
Deps: Open Source Insights project from Google Graph for Kubernetes 1.0.0
Graph for Kubernetes 1.22.0-alpha.2
Verifiable Supply Chain Metadata with Tekton Chains
Kubernetes CVEs: CVE-2021-25736
CVE-2021-25737
CVE-2021-25738
runc CVE-2021-30465
VS Code Plugin for Kubernetes CVE-2021-31938
Steve Smith says "GitOps is a placebo" in a blog post and Twitter thread Follow up from Vic Iglesias
GitOpsDays
Styra raises $40m Series B round Episode 101, with Tim Hinrichs and Torin Sandall
Cloud Native community goes live with 10 shows on something called Twitch
YouTube playlist for KubeCon EU 2021
Links from the interview Episode 92, with Pramod Ramarao
Dogecoin
Training and inference
12 things that prove Doom will run on literally anything
"It runs Doom" subreddit
CUDA
vGPUs
Multi-Instance GPUs
GKE now supports multi-instance GPUs
7 core MacBook Air GPUs
A100 GPU
16 A100 GPUs on a Google Cloud VM
Running GPUs on GKE Node taints for scheduling
NVIDIA Container Toolkit
GCP NVIDIA GPU device plugin
Kubernetes NVIDIA device plugin
GTC 2021 talks: A Deep Dive on Supporting Multi-Instance GPUs in Containers and Kubernetes by Kevin and Pradeep
Gain Competitive Advantage using ML Ops: Kubeflow and NVIDIA Merlin and Google Cloud by Andrew Stein and Maulin Patel (Google) and Davide Onofrio (NVIDIA)
Kevin's KubeCon talk and slides
Kevin Klues on Twitter
