Research · Preprint · 2026-09-05

Deterministic Integer Topologies

Bypassing the Matrix-Multiplication Bottleneck via Cache-Resident Graph Traversal

Matt Gibson · Crimson Symphony Media · Apache 2.0 companions on GitHub / HF

Cite
Concept DOI (stable cite): 10.5281/zenodo.22406009
Latest version: 10.5281/zenodo.22431987
Zenodo latest: zenodo.org/records/22431987
PDF (this site): zero-gemm.pdf · zero-gemm.md (machine ingest)
ORCID: 0009-0001-4167-201X
Code: github.com/ultranetcommand-neo/t112-l2-lattice
Dataset: huggingface.co/datasets/UltranetCommand/t112-l2-lattice

Figures

74.16 KiB T112 lattice occupying 28.9 percent of 256 KiB Kaby Lake L2 versus dense GEMM that cannot halt — Fig. 1 — Same table, two sockets. 74.16 KiB is 28.9% of a 2017 256 KiB L2. One 1024×1024 float32 GEMM on the 7700 is 767,000 hops.

KV cache compression still bills per token; Zero-GEMM is a 74.16 KiB total LUT with a hard halt — Fig. 2 — Industry 2026 compresses transformer KV (bytes × tokens × layers). This paper does not keep a KV. The working set is the whole engine.

What this is

A public receipt for cache-resident graph traversal on the T112 L2 lattice — integer topology work that aims at the GEMM bottleneck without claiming a finished OS, a live node count, or a retail install path.

Not MASTER. Not retail. Not a live inventory. Companion code and LUT telemetry are for reproduce-and-check.

Concept figure

Related figures

NIST ASD: Ne I near 6328 Å in air; T112 = 6328 as count, not a claimed spectral identity

Nature capture only: He–Ne 632.8 nm = 6328 Å, neon line. Pre-existent. Count ≠ wavelength. Gravity not claimed. No sermon on this page.

Related public anchors

Metric Tensor Folding DOI 22078565 proof.md — cos θ = 1/3 checksum.txt llms.txt

Page prepared 2026-09-05 for Pages bundle. Live only after operator yes on Direct Upload.