Open source FPGA NVMe accelerator platform for BPF driven ML processing with Linux/...
summary
An open source FPGA NVMe accelerator platform that runs Cortex‑A53 + Cortex‑R5 with Linux and Zephyr uses remoteproc/OpenAMP to pass custom NVMe commands and run a microBPF VM; NVMe register and FPGA Chisel code are auto‑generated from the NVMe spec PDF. TensorFlow Lite delegates call a VTA accelerator (from TVM) so BPF bytecode can offload ML kernels to FPGA. Code is on Micro’s GitHub under Apache‑2.0 and the prototype successfully executes delegated ML pipelines, enabling in‑drive processing and on‑the‑fly anonymization.
AI-written summary · about
chapters (11)
- 0:00Introduction and speaker background
- 1:02Why computational storage accelerators
- 3:52Hardware overview and platform options
- 4:35System architecture inside FPGA
- 7:46Dynamic code generation for FPGA and software
- 11:20Handling NVMe commands in Linux and Zephyr
- 13:29BPF virtual machine and ML integration
- 17:22Accelerator integration with VTA
- 21:55Open source availability and setup
- 23:21Research platform and future outlook
- 24:04Q&A session
related talks
34mBuilding FPGA-based Machine Learning Accelerators in Python
31mCFU Playground
16mSoftware-driven Hardware in the Data Center: LibreBMC, Renode, RowHammer...
39mSoftware Defined Hardware on Heterogeneous SoCs Using OpenAMP
36mTinyML at the Edge: Deploying and Optimizing AI Workloads on Zephyr RTOS
34mZephyr and OpenAMP
01h 19mMini-conference: RISC-V Support in Zephyr
10mDemo: NEOX TM : An Ultra-Low Power RISC-V GPGPU Enabling 3D & TinyML