Speaker
Description
Abstract
UALink is a new open, scale‑up interconnect standard for AI that tightly couples CPU, GPU, and NPU memory into a single high‑bandwidth, low‑latency domain, enabling ultra‑fast load/store/atomic operations across hundreds to thousands of accelerators in a pod. Unlike vendor‑specific fabrics such as NVIDIA's NVLink, UALink is positioned as a multi‑vendor standard for open accelerator interconnects.
This session proposes the bring‑up of a dedicated drivers/ualink subsystem and a minimal viable roadmap towards an upstream‑first path for new upcoming UALink hardware (Q3 2026).
Key points we'll explore:
- Vendor‑agnostic dma_buf memory semantics: Defining tightly coupled memory mappings in a vendor‑neutral way via
dma_buf, including pinning, revocation, and peer access across heterogeneous accelerators. - UALink daemon for fabric‑wide coordination: Introducing a UALink userspace daemon to coordinate UALink nodes across a rack, enabling cross‑node discovery, topology management, and policy enforcement.
- Security and confidential computing: Positioning UALink security as a first‑class part of confidential computing, protecting multi‑tenant and multi‑OS boundaries while still allowing high‑performance accelerator‑to‑accelerator data flow.
- RAS and port migration: Treating port migration and related scenarios as part of comprehensive RAS handling, and identifying where UALink can reuse kernel components from CXL (enumeration, RAS, error handling) versus where fabric‑scale accelerator memory requires new abstractions.
Goal: Reach consensus on minimal kernel interfaces to make UALink fabrics a first‑class Linux interconnect, complementing CXL while offering an open alternative to NVIDIA's proprietary NVLink and a clear standout from Broadcom SUE in delivering a multi‑vendor, fabric‑scale accelerator interconnect.
While awaiting new UALink‑based devices by Q3 2026, we propose a QEMU‑based UALink hardware simulation with a test DRM memory device to create a fully functional proof‑of‑concept of the proposed UALink core subsystem.
Tagline
Bringing UALink into Linux: defining dma_buf semantics and secure CPU/GPU/NPU memory sharing, positioning an open multi‑vendor alternative to NVIDIA's proprietary NVLink that complements CXL and clearly differentiates itself from Broadcom SUE.