Open Source NVMe AI Accelerator Platform with Zephyr
summary
An open source NVMe AI accelerator platform actually exists: Zephyr on Cortex‑R handling NVMe registers and DMA, an UltraScale+ FPGA fabric hosting a TVM/VTA accelerator, and openAMP bridging to Linux on Cortex‑A so unknown commands get forwarded. The system accepts custom NVMe commands to upload uBPF payloads and TensorFlow Lite models, runs ML pipelines in‑place on stored data, and proves the concept — performance relies on the FPGA accelerator. Code release goes to Chips Alliance GitHub soon.
AI-written summary · about
chapters (19)
- 0:00Intro
- 0:22Background Info
- 1:36NVMe Definition
- 3:09Why Accelerators in Hard Drives
- 6:10Open Source Platform
- 7:45Dedicated Platform
- 8:05Development Board
- 8:42Platform Overview
- 9:51NVMe Commands
- 10:50How does NVMe work
- 12:29Why Zephyr
- 13:30How to handle changes
- 15:06Configurability
- 19:16Open AMP
- 22:18Tensorflow Lite
- 23:36Virtual Accelerator
- 25:20What next
- 26:08Summary
- 27:04Questions
related talks
18mThe AI-First Data Infrastructure
57mPervasive and Sustainable AI with Adaptive Computing
23mRunning an Open Source LLM on Open Source Hardware
12mKeynote: Build Accelerated LLM Applications with Open Source Tools from NVIDIA
15mThe Open Ecosystem for the Agentic Age
13mKeynote: To Nemotron and Beyond: Building the Open Source AI Ecosystem
15mBuilding Gen-AI experiences using LLM and LVM
29mOpen Source Sovereign AI: Building Enterprise GenAI Platforms for Control & Comp...