How can artificial intelligence performance be improved through programming languages, compilers and advanced optimization techniques?
This lecture presents contemporary approaches to optimizing AI systems, with a special focus on operator fusion — a technique that enables multiple machine learning operations to be combined into a single, more efficient unit. In this way, data reuse is improved and the need to transfer data from global memory is reduced, which is particularly important for computation on graphics processing units.
The lecture will focus on Neptune, a tensor compiler developed for advanced operator fusion in complex sequences of reductions, such as the attention mechanisms used in contemporary artificial intelligence models. Unlike existing approaches, Neptune introduces a new method of optimization that enables more efficient computation and improved performance across different GPU architectures.
The lecture will demonstrate how Neptune, starting from ordinary code for variants of attention mechanisms, can generate operators equivalent to manually written FlashAttention and FlashDecoding operators. The results show that Neptune outperforms existing compilers and implementations, including Triton, TVM, FlexAttention and Triton-based FlashAttention implementations. Significant speedups have been achieved across different NVIDIA and AMD GPU architectures, demonstrating the potential of this approach for demanding deep learning workloads.
Neptune was presented at the PLDI’26 conference and is available via the DOI link .
At the end of the lecture, opportunities for collaboration through UIUC summer internships and doctoral dissertation work will also be presented.
The lecture will be delivered by Dr Saša Misailović, Associate Professor at the University of Illinois Urbana-Champaign – UIUC, where he has worked since completing his doctoral studies at MIT.
His work spans programming languages, compilers and software engineering, with a focus on programming systems for applications in which noise in data, unreliable communication, approximate computing and uncertainty are integral parts of computation. Such applications are present in numerous fields, including artificial intelligence, data analytics, multimedia processing and robotics.
Recently, the main focus of his research group has been the development of innovative programming systems that make AI applications more accurate, reliable, robust and resource-efficient. His work connects compiler infrastructure for complex AI systems, reliable constrained generation for large language models, robustness analysis of AI programs, as well as the synthesis and analysis of software that operates under uncertain conditions.
Admission is free.