Architecting Cloud Storage for AI Native Applications
summary
Vultr presents an AI-native cloud storage stack with 32 global data centers and >60 PB of Ceph-backed storage deployed, mixing S3-compatible object, resizable block tiers (HDD/SSD/NVMe) and a new Vultr File Systems read-write-many that attaches to multiple instances for training and inference. Production clusters use AMD MI300X nodes (8 GPUs per OAM) delivering ~50 GB/s per GPU over RoCE/InfiniBand; the stack is live and tuned for distributed training and low-latency regional inference.
AI-written summary · about
chapters (9)
- 0:00Introduction and welcome
- 1:46Cloud infrastructure evolution
- 3:12AI native application architecture
- 5:18AMD MI300X GPU cluster storage
- 7:49Data governance and storage strategy
- 8:11Vulture file systems announcement
- 10:57High-speed networking for AI workloads
- 12:42Storage requirements for AI training
- 13:06Q&A and closing remarks
related talks
22mScaling agentic AI: From infrastructure to enterprise outcomes – Duncan Ng
31mAI, ML, and the Ceph Advantage: Scalable Storage for Smarter Workflows
05mFrom Open Source to Real-World Scale
36mCloud Native Storage CubeFS, Empowering AI Acceleration
24mHow to Eliminate the I-O Bottleneck & Continuously Feed the GPU While Training in the Cloud
25mOptimize productivity by meeting next-level demands of the AI era
04mA Cloud Native Stack From Bare Metal To Tokens for Large-Scale AI Inference
29mAccelerate your cloud architectures with the right storage options