AMD has entered the AI arena with its first fully open large language model (LLM), called OLMo. Similar to OpenAI’s GPT-4, OLMo is designed to handle complex reasoning and conversation tasks. The model, available for download, can be run on AMD Instinct MI250 GPUs and PCs equipped with Ryzen chips featuring neural processing units (NPUs). By offering this model through its AMD Developer Cloud, the company is positioning itself as a key player for both data centers and smaller businesses seeking AI solutions.
OLMo’s flexibility allows businesses without high-end hardware to leverage its capabilities using PCs powered by Ryzen AI. This accessibility could attract organizations looking to integrate AI into their workflows without the need for specialized infrastructure. Additionally, the open-source nature of OLMo means that developers can fine-tune the model with their own datasets, creating custom AI tools tailored to specific needs.
The OLMo series was trained using AMD’s Instinct MI250 GPU clusters, involving 1.3 trillion tokens across 16 nodes, each equipped with four MI250 GPUs. The model went through three phases: pretraining, supervised fine-tuning (SFT), and direct preference optimization (DPO). These stages enhanced the model’s performance in areas like science, coding, and mathematics while aligning its responses to human values.
In benchmarking, OLMo performed strongly, particularly in evaluations like AlpacaEval2 and MT-Bench, where it outperformed other open-source LLMs. It also matched its peers in responsible AI tests, such as TruthfulQA and ToxiGen, showing promise as a trustworthy and non-toxic conversational agent.
AMD’s OLMo is now available for free, marking a significant step forward in the democratization of AI technology, enabling both large and small organizations to innovate and develop AI-driven solutions.
