Workshop: IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms
Authors: Milan Shah (North Carolina State University), Sheng Di (Argonne National Laboratory), and Michela Becchi (North Carolina State University)
Abstract: In recent years, novel AI accelerators have emerged as promising alternatives to GPUs for AI model training and inference. The Cerebras CS-3 has demonstrated strong performance on machine learning and some scientific applications. While the benefits of Cerebras systems for dense workloads are well demonstrated, their potential for sparse workloads, particularly for large matrices with substantial memory requirements, is not yet understood.
In this work, we study two sparse linear algebra kernels, sparse-dense matrix multiplication (SpMM) and sampled dense-dense matrix multiplication (SDDMM), on the CS-3. We propose low-level CS-3 designs and optimize them for I/O performance, memory footprint, and scalability. Our evaluation shows that the CS-3 kernels can outperform the CPU by up to 100x for SpMM and 20x for SDDMM on 90% sparse matrices, with SpMM performance improving as matrix dimensionality increases. However, beyond 99% sparsity, our kernels suffer performance degradation and approach or underperform state-of-the-art CPU libraries.
Back to IA^3 2026 — 16th Workshop on Irregular Applications: Architectures and Algorithms Archive Listing
Back to Full Workshop Archive Listing