Architecture-General Non-Bijunctive Hebbian Collapse for LLM Inference: From PowerPC vec_perm to Apple Silicon NEON
Abstract
We demonstrate that non-bijunctive permutation collapse — amplifying strong attention paths and pruning weak ones via hardware byte-shuffle instructions — is architecture-general. Originally developed on IBM POWER8 using vec_perm, we port to Apple Silicon NEON (vqtbl2q_u8) and show identical behavioral properties: 91% attention sparsity, consistent persona-level divergence, zero inference overhead. We introduce CPU-guided sparse FFN via Metal unified memory, enabling real-time pruning of 60-70% of feed-forward computation — impossible on discrete GPU architectures.
Fields of Study & Subjects
Parallel Computing and Optimization Techniques
Advanced Neural Network Applications
Topic Modeling
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 16 March 2026
- Language
- EN
- Document type
- Preprint