AMD posts AIE driver for Versal AI Engine upstream Linux kernel review
AMD submitted a new open-source driver for the Versal AI Engine accelerator to the mainline Linux kernel, using the standard accel framework rather than a downstream proprietary solution.
Summary
AMD has posted patches for a new AMD AI Engine (AIE) driver to the dri-devel mailing list for mainline Linux kernel review. This driver targets the AI Engine found in Versal adaptive SoCs, a 2D array of VLIW SIMD vector processors used for DSP, machine learning, and other workloads. The submission is distinct from the existing AMDXDNA driver that supports Ryzen AI NPUs, and represents a reworked effort focused specifically on newer Versal Gen 2 SoCs.
Each Versal system can contain tens to hundreds of AI Engines, and until now no mainline kernel driver existed for this hardware. The new driver complies with the standard Linux accelerator (accel) subsystem framework, marking a shift from AMD's previous out-of-tree approach. This upstreaming effort follows an earlier, more basic patch series from the previous year.
Driver architecture and Versal hardware target
- Hardware target: The driver supports the AMD AI Engine on Versal adaptive SoCs, specifically targeting Versal Gen 2 devices rather than earlier generations.
- Processor architecture: The AIE consists of a 2D array of VLIW SIMD vector processors designed for DSP, machine learning, and related compute workloads.
- Scale: Versal SoCs can deploy between tens and hundreds of AI Engines per individual system or device.
- Kernel framework: The driver implements the standard Linux kernel accelerator (accel) subsystem, replacing AMD's previous downstream out-of-tree driver approach.
- Distinction from AMDXDNA: This AIE driver is separate from the existing AMDXDNA driver already in mainline Linux, which handles Ryzen AI NPUs rather than Versal SoCs.
Why it matters
Upstreaming Versal AI Engine support through a standard kernel framework reduces fragmentation for embedded and edge Linux distributions targeting AMD adaptive SoCs, and enables long-term maintenance without proprietary driver dependencies. For developers working with FPGA-accelerated ML and signal processing pipelines, this provides a predictable interface that aligns with other Linux accelerator drivers rather than a vendor-specific stack.
Original reference: www.phoronix.com →