From Model to Metal: An Open Source Stack for Accelerating Intelligence

Artificial intelligence is becoming truly ubiquitous: from agentic AI systems in enterprise settings to everyday smart devices. Arm’s vision is to enable this intelligence to run efficiently everywhere. This talk outlines the technical architecture needed to support advanced AI on Arm CPUs and accelerators. It will highlight the strategic role of open source tools in unifying the currently fragmented AI software landscape. At the core is a PyTorch model for which training and inference is accelerated with tools such as vLLM, Ray and oneAPI that optimize everything from matrix operations to data pipelines and distributed workloads. Ultimately, this open ecosystem approach unlocks faster innovation, greater developer velocity, and reliable deployment at scale and ensures that AI stack evolves in sync with the hardware. Open source collaboration allows the industry to pool its efforts on common tooling, avoiding duplicated work and ensuring that breakthroughs benefit everyone.