Persistent State Machines for LLM Attention (Vivado Estimated 0.0267 PJ/Op)

yusuke_esaka1 pts0 comments

Persistent State Machine: A Formal Computational Paradigm for High-Sparsity LLM Attention Acceleration [Version 7.0] | Zenodo

Skip to main

You are using an outdated browser. Please upgrade your browser to improve your experience.

Published July 30, 2026

| Version v7

Preprint

Open

Persistent State Machine: A Formal Computational Paradigm for High-Sparsity LLM Attention Acceleration [Version 7.0]

Authors/Creators

Esaka, Yusuke1

Show affiliations

1.

Cosmos Administrative Scrivener Office & Independent Researcher

Description

Persistent State Machines: Complete Mathematical Proofs and Vivado Implementation Synthesis (Version 7.0)

ABSTRACT:

We present a formal discrete framework for attention operators in Large Language Models (LLMs) via Persistent State Machines (PSMs). Computation is broadcast as instructions to stationary in-memory cells that evaluate local deterministic state transitions.

Complete mathematical proofs are provided for quantization error bounds, a concrete multi-phase discrete Softmax construction under an explicit bounded-logits assumption, deterministic finite-automaton (DFA) equivalence with spatial factorization into O(N) circuit size, and membership in DSPACE(O(n)).

A full two-phase engine—comprising local score evaluation and binary-tree reduction (including global max extraction and local exponential lookup)—was implemented in synthesizable RTL and processed through the AMD Vivado 2026.1 tool flow (logic synthesis, placement, routing, static timing analysis, and post-implementation power estimation) targeting the Zynq-7000 xc7z020 device.

After correcting a previous throughput exponent error in earlier drafts, the tool-estimated normalized dynamic energy of the synthesized logic is 0.0267 pJ/op under the stated operating conditions (Fmax = 283.8 MHz). Functional simulation with over one thousand random test vectors confirmed bit-exact agreement (zero discrepancy) with a fixed-point Python software reference.

Note: All reported energy and timing figures are tool estimates produced by the Vivado flow for synthesized logic; no physical FPGA board execution or silicon measurement was performed, and external system-level memory energy is excluded.

Japanese Patent Application No. 2026-177318 (Patent Pending)

Files

026_Zenodo_ASMA_Paper_v7_0_FullProofMaster.pdf

Files<br>(74.3 kB)

Name<br>Size

Download all

026_Zenodo_ASMA_Paper_v7_0_FullProofMaster.pdf

md5:eec884040073d8a06b69078645cb6d1b

74.3 kB

Preview

Download

1K

Views

35

Downloads

Show more details

All versions<br>This version

Views

Total views

1,056

Downloads

Total downloads

35

Data volume

Total data volume

1.4 MB<br>0 Bytes

More info on how stats are collected....

Versions

External resources

Indexed in

OpenAIRE

Communities

Details

DOI

DOI Badge

DOI

10.5281/zenodo.21694686

Markdown

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21694686.svg)](https://doi.org/10.5281/zenodo.21694686)

reStructuredText

.. image:: https://zenodo.org/badge/DOI/10.5281/zenodo.21694686.svg<br>:target: https://doi.org/10.5281/zenodo.21694686

HTML

Image URL

https://zenodo.org/badge/DOI/10.5281/zenodo.21694686.svg

Target URL

https://doi.org/10.5281/zenodo.21694686

Resource type<br>Preprint

Publisher<br>Zenodo

Rights

License

Creative Commons Attribution 4.0 International

The Creative Commons Attribution license allows re-distribution and re-use of a licensed work on the condition that the creator is appropriately credited.

Read more

Citation

Export

Technical metadata

Created

July 30, 2026

Modified

July 30, 2026

Jump up

This site uses cookies. Find out more on how we use cookies

Accept all cookies<br>Accept only essential cookies

zenodo state https persistent version attention

Related Articles