Day 1 Keynote
summary
Composable GPU fabric that lets servers attach or release GPUs live over PCIe channel based GPU provisioning, built on a Power9 OpenPOWER switch and a 16‑GPU chassis, claims sub‑network latency and multi‑GB/s per‑link bandwidth using P2P over PAM‑4/PCIe fabric. GPUs bind to any host without reboot, expose real‑time memory and utilization counters, and raise measurable GPU iteration/utilization rates versus fixed servers. A working demo boots and shows dynamic allocation; team reports lower TCO for pooled GPU deployments.
AI-written summary · about
related talks
18mArchitecting Multi-tenant Data-center Networks for GPU Customers
39mFunction Dedicated Network Deployment Model in DPU/IPU-Like Device Acro...
29mThe Path to GPU as a Service in Kubernetes
13mDemocratizing AI with Open Source Compute
16mBOSH & GPU - a Match Made in Heaven
14mWhy We Don’t Need More Data Centers
15mTransparent GPU Sharing in Container Clouds for Deep Learning Workloads
09mDemocratizing Access to GPUs for Faster MLAI Innovation with Generative AI