RPI: Resonant Permutation Inference - Zero-Multiply Language Generation via Cache-Resonant Permutation Routing
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
RPI is a language generation engine that eliminates all floating-point multiply-accumulate operations from inference. It replaces matrix multiplication with permutation table lookups and integer accumulation, achieving 18,000 tokens/second with model sizes of 0.3-3 MB. Introduces cache-resonance attention (using CPU cache hierarchy as implicit attention), NUMA-aware cell banking, and a dual-brain architecture for speculative LLM drafting. Runs on any architecture with a CPU cache hierarchy: ARM (NEON tbl), x86/x86_64 (SSE/AVX), PowerPC G4 (AltiVec vperm), PowerPC G5, IBM POWER8 (VSX vec_perm), MIPS (N64 R4300i), AArch64, RISC-V, SPARC, and vintage x86 (386+). Theoretically any CPU that can do table lookup and integer add/subtract.
Fields of Study & Subjects
Parallel Computing and Optimization Techniques
Natural Language Processing Techniques
Topic Modeling
Publication details
- Journal / source
- Zenodo (CERN European Organization for Nuclear Research)
- Publication date
- 28 March 2026
- Language
- EN