Large language models and other machine learning methods to advance AMD’s hardware and software capabilities

AMD is a leading innovator in high-performance computing, graphics, and visualization technologies, focusing on gaming, immersive platforms, and data center solutions. The company develops cutting-edge hardware
and software solutions to enhance computing performance across various applications, including artificial intelligence (AI), machine learning (ML), and edge computing. AMD is seeking to apply advances in AI and ML techniques, especially Large Language Models (LLMs) to address various technical development opportunities from internal process innovation to improving its software and hardware capabilities in gaming and video processing capabilities. The research areas AMD are seeking to address in the proposed project falls into the following research themes. Research Themes
1. Code optimization for asynchronous and synchronous parallel execution across its diverse array of processor types (such as CPUs and GPUs).
2. With the rapid advancement of LLMs, deploying state-of-the-art AI models presents significant memory and compute capacity challenges, especially on embedded and client platforms. While cloud-based AI inference is widely used, latency, privacy, and cost constraints make edge-based AI processing increasingly critical. AMD aims to optimize heterogeneous inference for LLMs by efficiently distributing workloads across CPUs, GPUs, and NPUs within its APUs. Increasing difficulty in optimizing code for its diverse array of processor types (such CPUs and GPUs), which presents significant challenges in code optimization for asynchronous and synchronous parallel execution. Deploying LLMs with long context prompts also poses significant challenges in terms of memory and compute. This issue is critical in applications like multimodal LLMs, which involve massive input token sizes.
3. AMD’s strategic goal of enhancing AI capabilities for game development, specifically focusing on intelligent and adaptive non-player character (NPC) behaviors. Traditionally, crafting engaging and realistic non-player character behaviour (NPC) requires extensive manual effort. This project leverages recent advancements in Large Language Models (LLMs) and reinforcement learning (RL) to automate NPC training and scenario creation, significantly reducing development costs and timelines while improving game realism and player engagement through AMD Schola.
4. The User Experience Group at AMD Canada is developing a local chatbot for answering customer questions about AMD’s products. In order to ensure that the chatbot accurately communicates product information, the chatbot must use Retrieval Augmented Generation (RAG) to ground its responses in product data. However, traditional methods that use text chunking and vector embeddings on structured and unstructured data struggle to find dynamic relationships between text chunks and extract context for complex queries. This project aims to explore innovative techniques to extract complex relationships and insights from AMD’s knowledge base to enhance chatbot intelligence. Doing so will improve customer satisfaction by allowing the chatbot to more accurately answer customer questions, potentially increasing sales and decreasing the need for customers to communicate with customer service.
5. In the video processing domain, AMD aims to enhance real-time video upscaling capabilities to improve user experience during video playback and streaming. Although existing solutions like Radeon Super Resolution and FidelityFX Super Resolution (FSR) are present, challenges persist in achieving highquality upscaling and super-resolution for compressed video content without introducing artifacts or latency. In addition, conventional high dynamic range (HDR) imaging, which merges multiple standard dynamic range (SDR) images, is impractical due to high computational costs, increased latency, power consumption, and motion artifacts. Meanwhile, the rising demand for video content necessitates more efficient processing techniques, as existing solutions fail to meet bandwidth requirements.
Successfully developing methods and solutions to the above research themes will advance the hardware and software capabilities in video processing and game development. The successful outcomes will also increase the general efficiency in deploying LLM with respect to memory usage and compute further increasing the capabilities of AMD’s hardware for video processing and game development. The economic benefits for AMD go beyond
gaining operational efficiency and promises broader impacts to the accessibility of high performance computing (HPC) by simplifying the process for creating optimized code for power gaming, immersive platforms and data centres, as well as retaining its leadership position in the semiconductor industry.

Faculty Supervisor:

Marsha Chechik;Steve Engels;Igor Gilitschenski;Natalie Enright Jerger;Angela Demke Brown;Aviad Levis;Scott Sanner;Gerald Penn;Florian Shkurti

Student:

Partner:

AMD Canada

Discipline:

Computer science

Sector:

Manufacturing; Professional, scientific and technical services

University:

University of Toronto

Program: