Microsoft open-sources DACS Kubernetes caching system

Microsoft has open-sourced the distributed accelerator for cloud storage (DACS), a Kubernetes-native caching system designed to accelerate AI inference and training workloads. By leveraging local NVMe or SSD storage in GPU cluster nodes, DACS serves repeated reads for large artifacts like model weights and datasets directly from cache while using Azure Blob Storage as the durable source of truth. Initial evaluations show that DACS significantly reduces dataset-loading delays and mitigates inference startup throttling caused by concurrent blob fetches. Open article →