Uploaded July 2026 | Updated September 2026, 2 weeks ago

---

Abstract: KV cache is content-addressed by nature — so why are we serving it over block, with lookup tables mapping hashes to offsets? This talk introduces rados-nkv, an open-source prototype that exposes Ceph's RADOS via the NVMe key-value command set. Using ROCm XIO, SPDK, and a QEMU MMIO bridge, an AMD GPU issues KV Retrieves from device code and receives values from a distributed Ceph cluster DMA'd straight into VRAM. One doorbell ring fetches a 262k-token prefix. We'll cover the architecture and start talking about what's next.
Featuring:

Kyle Bader, Chief Architect, Data & AI, Ceph at IBM
Stephen Bates, Fellow at AMD

Slides: ceph.io/assets/pdfs/tech-talks/rados-nkv-tech-talk.pdf
Blog post: ceph.io/en/news/blog/2026/for-whom-the-door-bell-tolls


Submit a speaker request here: airtable.com/apphc2dbSP8GuCdor/shrm3t1Fwue5YfAlf.

See upcoming talk here: ceph.io/en/community/tech-talks

Watch previous talks on our YouTube playlist: youtube.com/playlist?list=PLrBUGiINAakM36YJiTT0qYepZTVncFDdc.