Computer Architecture Research

Impact: HipKittens addresses the critical software gap limiting AMD GPU adoption in AI workloads, often called the 'CUDA moat.' By providing accessible C++ programming primitives, it enables developers to write high-performance AMD kernels without resorting to raw assembly. The framework achieves 1.2-10× speedups over existing baselines in various settings and matches AMD's hand-optimized assembly kernels across key operations like GEMM and attention. This work is particularly impactful for democratizing AI hardware access, as AMD MI355X GPUs offer competitive or superior specifications to NVIDIA alternatives (2.5 PFLOPs BF16, 8 TB/s bandwidth, 288 GB memory). The open-source release enables the AI community to leverage diverse hardware platforms, potentially breaking vendor lock-in and accelerating AI development through increased compute availability.

Impact: Nested Learning provides a new theoretical framework for understanding and designing machine learning models with enhanced continual learning capabilities. The HOPE architecture demonstrates practical improvements over Transformers and modern recurrent networks in language modeling tasks, achieving better perplexity and accuracy on benchmark datasets. The framework addresses the static nature of Large Language Models after deployment by enabling online memory consolidation, similar to human brain processes. This has significant implications for developing AI systems that can continually acquire new knowledge beyond their immediate context window, potentially reducing the need for expensive retraining and enabling more adaptive AI systems in production environments.

Impact: PAN enables practical applications in robotics, autonomous systems, and AI planning by providing a general-purpose simulator that can predict future world states based on natural language actions. The model's ability to maintain long-horizon consistency and support interactive simulation makes it valuable for training robotic policies, testing autonomous vehicle scenarios, generating synthetic training data, and enabling agents to perform 'thought experiments' before taking real-world actions. Its open-domain generalization allows deployment across diverse environments without domain-specific retraining, while the natural language action interface makes it accessible for human-in-the-loop applications and planning systems.

Impact: Directly reduces time-to-first-token (TTFT) in production LLM serving systems, enabling better user experience for long-context applications like RAG, summarization, and in-context learning.

Impact: Directly applicable to next-generation GPU architectures, with major vendors expressing interest in the approach for future products.

Impact: Enables cost-effective scaling to 256+ cores without prohibitive directory memory overhead, directly applicable to datacenter processors.

Impact: The Transformer architecture has revolutionized natural language processing and beyond, becoming the foundation for modern large language models like BERT, GPT, and their successors. Its parallel processing capabilities enable efficient training on modern GPU hardware, reducing computational costs and training time significantly. The architecture's success in machine translation demonstrated that attention mechanisms alone could outperform recurrent networks, leading to widespread adoption across various domains including computer vision, speech recognition, and protein folding. The model's interpretability through attention visualizations has also provided insights into how neural networks process sequential data, influencing both research and production systems in industry.