• Investigating the latest and future NVIDIA GPU hardware architecture and programming models • Working on the frontier of AI by understanding advanced algorithms (like attention sinks and MoEs) and numerics (like block-scaled floating point) to identify new opportunities for optimization • Designing and implementing compiler technology using MLIR to optimize high-level kernel descriptions (written in Triton's Python DSL), with a focus on generating efficient, low-level GPU code • Engaging in a dynamic, iterative process of optimization—sometimes starting with the kernel, sometimes with the compiler—to find the most efficient path to peak performance • Collaborating with teams across NVIDIA, including hardware architects and the CUDA compiler team, to influence future products and ensure we are always operating at maximum efficiency