Choose timezone
Your profile timezone:
NGT Tutorials: AIE4ML
→
Europe/Zurich
40/S2-B01 - Salle Bohr (CERN)
Description
AMD AI engines are specialized processors specifically designed to run AI inference in high performance, but they are also difficult to program in an efficient manner due to the numerous compute units, custom data paths, and local memory. For example, the AMD’s latest AI Engine (AIE-ML) has a grid of vector processors and associated memory tiles that can compute layers of a neural network entirely on chip, which is highly important for low-latency tasks.
This tutorial introduces aie4ml, a compiler-style backend that targets machine-learning models for execution on AMD AI Engine devices, currently supporting AIE-ML and AIE-MLv2. The aie4ml compiler takes entire neural networks and compiles them down to highly efficient, hardware-aware code, fully exploiting on-chip memory and compute while maintaining bit-exact results with low-precision (quantized) models. This tutorial takes you from first principles to hands-on usage; you’ll see how a neural network model is taken from a familiar training framework, transformed through aie4ml’s toolflow, and turned into an optimized project that can be simulated or built for actual hardware. You’ll explore key concepts such as how the framework maps operations across the 2D array of AI engines, manages data placement and communication, and ultimately achieve GPU-class throughput. By the end, you’ll understand not only what aie4ml does behind the scenes but also how to use it to convert, inspect, and run your own models on Versal AI engine platforms.
This tutorial introduces aie4ml, a compiler-style backend that targets machine-learning models for execution on AMD AI Engine devices, currently supporting AIE-ML and AIE-MLv2. The aie4ml compiler takes entire neural networks and compiles them down to highly efficient, hardware-aware code, fully exploiting on-chip memory and compute while maintaining bit-exact results with low-precision (quantized) models. This tutorial takes you from first principles to hands-on usage; you’ll see how a neural network model is taken from a familiar training framework, transformed through aie4ml’s toolflow, and turned into an optimized project that can be simulated or built for actual hardware. You’ll explore key concepts such as how the framework maps operations across the 2D array of AI engines, manages data placement and communication, and ultimately achieve GPU-class throughput. By the end, you’ll understand not only what aie4ml does behind the scenes but also how to use it to convert, inspect, and run your own models on Versal AI engine platforms.
NGT Tutorials