Speaker: Dr. Ahmed Mahmoud
Date: Thursday, March 19, 2026
Time: 11:00 am to noon
Location: online
Meeting ID: 946 3073 6303
Passcode: AHMED
Abstract:
The last decade of hardware scaling has been defined by a contract of regularity in which modern parallel hardware reaches peak throughput when data is organized as dense tensors and computation is uniform. This contract breaks down for the irregular representations used to model the physical world, including dynamic meshes in simulation, sparse matrices in scientific computing, and graphs in network analysis. In these workloads, we realize only a fraction of peak utilization because irregular memory access limits bandwidth efficiency and nonuniform execution fragments parallelism across the machine. In this talk, I argue that irregularity does not imply a lack of structure. I present a research agenda that bridges the gap between irregular computation and parallel hardware by exploiting hidden structures (such as geometric locality, sparsity, and low rank) that current systems ignore. I demonstrate a systems view that makes latent structure explicit and uses it end-to-end, from representation to kernels to applications. At the data structure level, I introduce RXMesh, a GPU system that leverages patch-based locality to enable high-performance dynamic topology changes entirely on the GPU, achieving up to 80× speedups over state-of-the-art frameworks. At the computational primitive level, I show how to exploit sparsity in automatic differentiation for efficient sparse Hessian and Jacobian assembly and how to leverage hierarchical separators to reduce fill-in and accelerate sparse Cholesky factorization. At the application level, I demonstrate how exploiting low-rank structures in neural fields can reduce memory footprints by 7–8×, enabling parameter-efficient representations for visual computing. Finally, I outline future directions for generalizing these principles beyond geometry, aiming to build native parallel systems for a broader class of irregular computational problems, including large-scale differentiable graph processing, numerical optimization, and randomized linear algebra.
Biography:
Ahmed Mahmoud is a postdoctoral associate at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) where he works with Professors Jonathan Ragan-Kelley and Justin Solomon. His research focuses on high-performance computing, scientific computing, and computer graphics with an emphasis on designing systems, algorithms, and data structures that accelerate complex, irregular problems. Prior to MIT, he served as a senior research scientist at Autodesk Research from 2021 to 2024. He earned his Ph.D. in Electrical and Computer Engineering from the University of California, Davis, in 2024 and received his M.Sc. from the same department in 2020. He holds a B.Sc. in marine engineering and naval architecture.