My master’s project: a GPU library for accelerating linear algebra operations on large batches of small matrices. Achieves up to ~90x speedup over state-of-the-art libraries such as MAGMA and JAX. Overview # Modern GPU-accelerated libraries are built for large matrices, where each operation carries enough work to saturate the hardware on its own. In many applications, though, the workload is instead a very large batch of small matrices, a pattern common in particle filters, Kalman methods, and parts of ML. Here, the existing libraries fail to fully utilise the GPU, and at sufficiently large batch sizes, runtime becomes prohibitively slow.