Work

Native GPU Octonion Support for Deep Learning: A Compiler-First Approach

Year: 2026
Type: preprint
Abstract: We present, to our knowledge, the rst compiler-level implementation of GPU-accelerated octonion algebra for deep learning. Octonionsthe largest normed division algebra over the realsoer 8× parameter eciency compared to real-valued networks while preserving the mathematical structure necessary for physics-informed machine learning. Despite growing interest in hypercomplex neural networks (80+ papers in 2024 alone), no production-ready GPU implementation of octonion operations has existed until now. Our implementation in the Sounio compiler provides 20 GPU-accelerated operations targeting both NVIDIA (PTX) and Apple Silicon (Metal), validated against the Moufang identities with all 72 mathematical and platform tests passing. We achieve 1416 GFLOPS throughput on Apple M3 Max and 811 GFLOPS on Intel Xeon, demonstrating cross-platform viability. This work addresses an explicit gap in the literature where researchers have called for parallelized GPU/TPU kernels for fast octonion products.
Language: English
Open Access status: green
OpenAlex ID: W7127069269
Cites:
Cited by:
Related to: 10