AutoPhi CS-51-AI — AI-Pipeline Preprocessing Storage Reference Kit
Publicly online since 2010 · U.S. patent applications since 2012 · inventions offered since 2014. The work of Christopher Gabriel Brown, independently documented.
Downloads immediately on checkout — the CD GDSII Generator (AES-256 encrypted). Open with 7-Zip, WinRAR, or macOS Archive Utility using password CRIONE99KEY. Founder Access Pass credit (code FAP99CREDIT) applies at checkout.
AutoPhi CS-51-AI — AI-Pipeline Preprocessing Storage Reference Kit
The AI-preprocessing vertex of the AutoPhi IC family. Five named AI-pipeline kernels — VECTOR_DOT, TOPK, ARGMAX, EMBED_LOOKUP, SOFTMAX — filed directly into the substrate and reduced near the data, so only the answer crosses the PCIe bus. Design target: 38× lower latency and 31,250× lower PCIe traffic than a traditional Gen5 NVMe SSD on a 1 GiB embedding-retrieval workload. In a RAG or vector-search pipeline, CS-51-AI runs the embedding-table fetch, vector scoring, and TOPK selection near the data — the H100 / B100 / B200 cluster downstream is not consumed by those preprocessing kernels and gets its cycles back for actual model inference. This is an offload story, not a replacement story.†
What you receive
- Full Mathematical Deposition contents — every claim, figure, and lineage doc anchored by USPTO application 19/710,460 (CS-50 substrate).
- Five named AI-pipeline kernels, filed into the substrate.
- VECTOR_DOT — signed integer inner product against a host query, evaluated near the data.
- TOPK — return the top-K records by a host-supplied scoring function, reduced through the on-die fabric.
- ARGMAX — return the LBA with the largest score across an arbitrarily large stripe.
- EMBED_LOOKUP — fetch fixed-prefix embeddings by global token ID. This kernel serves the twelve-point vector registration defined by the Hyper Quantum Map (Project 37, USPTO 19/646,681).
- SOFTMAX — numerically-stable softmax across a stripe, returned as a normalised probability vector.
- Reference Python simulator —
ace_scaleout.pywith the 18 / 18 PASS self-test suite covering all five kernels plus VOXEL / SCAN / XFORM at N = 1 byte-identical to CS-50 and N = 4 with the tree reducer. - Head-to-head benchmark —
benchmarks/ai_pipeline_demo.pyreproduces the 38× speedup and 31,250× PCIe-reduction headline numbers locally on the buyer’s laptop. The numbers are deterministic and reproducible from the spec rates — they will reproduce identically. - Architecture document — multi-die block diagrams + scale-out FTL + ACE fabric + tree reducer description, signed against
peer_fabric.py24 / 24 topology comparison. - Verify-it-yourself runner —
verify.pydrives all of the above against iverilog + vvp + Python and reports 12 / 12 PASS in under a minute. The buyer’s entire diligence is one command. - Bundle manifest tooling —
handoff/scripts/build_bundle.pywrites a SHA-256 manifest over the entire delivery;verify_bundle.pyre-validates it. 82 files hashed today. - Patent application references — full provenance for the five related USPTO filings, with PCT-window dates.
What it’s NOT
- NOT shipping silicon. CS-51 is a paper-tapeout reference kit. The buyer takes the foundry-handoff bundle to their own foundry account.
- NOT a CS-50 Evaluation or Full Use & Copy License. The Mathematical Deposition is a read-only deposition tier; full RTL evaluation and production rights are sold under separate AutoPhi family licenses.
- NOT a commodity AI-accelerator IP license. CS-51-AI is not priced or framed against Nvidia H100/B100/B200, AMD MI300, Hailo, Mythic, SambaNova, Cerebras, Tenstorrent, or Groq. AutoPhi is its own category; the deliverable carries kernel primitives filed directly into a computational-storage substrate, not an accelerator card.
- NOT a CS-51-PF replacement. CS-51-PF carries the peer fabric + VOXEL claim; CS-51-AI carries the kernel set. The Complete Scale-Out Bundle ($300M) carries both with a $45M discount.
- NOT exclusive. CS-51-AI may be licensed to other organisations at CRI-ONE’s sole discretion.
The refutable chain — every claim, every test count
iverilog_multichannel— 4 / 0 PASS — N-port FTL arbiter, host strict-priority + round-robin.iverilog_ace_scaleout— 5 / 0 PASS — N instances of CS-50’s ACE + tree reducer.iverilog_peer_arb— 2 / 0 PASS — primary/replica FSM, 128-bit framed peer fabric.iverilog_p2p_bridge— 6 / 0 PASS — 128-bit frame ↔ 256-bit MWr TLP encap/decap.ace_scaleout(Python) — 18 / 0 PASS — N = 1 byte-identical to CS-50; eight kernels including the five named AI-pipeline kernels above (VECTOR_DOT, TOPK, ARGMAX, EMBED_LOOKUP, SOFTMAX) plus VOXEL / SCAN / XFORM.peer_fabric(Python) — 24 / 0 PASS — five topologies, UCIE_INTRA_PKG wins at every N ≥ 2.ai_pipeline_demo— 38× latency, 31,250× PCIe-traffic reduction (design targets on a 1 GiB embedding-retrieval workload).tapeout_manifest— 1,128 files SHA-256 hashed (CS-50 frozen baseline).tapeout_cs51_manifest— 122 files SHA-256 hashed (CS-51 extension on top).compliance_forbidden— bundle markdown passes the prohibited-phrasing scan: no claims that overstate intellectual-property status, performance ratios, or third-party adoption.cross_references— every internal cross-reference resolves.placeholders_registered— 20 / 20 layers closed;--strict-shipGREEN.
† About the H100 reference. The Nvidia H100 SXM5 reference on this page is not a capability-replacement claim. An 8× H100 SXM5 baseboard provides general AI training and inference compute that CS-51 does not. The relationship described is offload: CS-51-AI runs the embedding-table fetch, vector scoring, and TOPK selection near the data, so the H100 cluster downstream is not consumed by those preprocessing kernels. The 38× latency and 31,250× PCIe-traffic-reduction numbers measure the embedding-retrieval slice in isolation and are reproducible via the included benchmarks/ai_pipeline_demo.py; they are not measurements against an 8× H100 baseboard.
Patent record
Why the headline price
The $95,000,000 anchor reflects CS-51-AI’s position as the AI-preprocessing vertex of the AutoPhi IC family, not a commodity AI-accelerator IP license. Comparable AutoPhi family members on cri-one.com/store carry these anchors: AutoPhi BGA-1536 (Project 44, production-ready KiCad sources) at $80M; ZettaFLOP 1Z Accelerator (Project 23) at $150K–$1.5B across three tiers; Software-Driven Data & SSD (Project 42) at $35M; Quantum Triplet π Framework (Project 37, the quantum cartography CS-51-AI EMBED_LOOKUP serves) at $5M–$50M; AutoPhi Quantum PCIe 5.0 (Project 35) at $5M–$25M; NeuroElement (Project 22) at $15M. CS-51-AI is positioned at the same magnitude as CS-50 Tier 3 Full Use & Copy because the five named kernels plus the substrate filing represent a complete near-data AI preprocessing substrate — one filed deposition that an AI infrastructure buyer uses to build an entire generation of inference and retrieval workloads on.
At fleet scale, the value is not a head-to-head replacement of an H100 baseboard — CS-51-AI doesn’t compete in general AI training or inference compute. The value is the GPU cycles CS-51 returns to a buyer’s H100 / B100 / B200 cluster by removing embedding-retrieval preprocessing from the GPU path entirely. How big that win is depends on what fraction of the buyer’s GPU cluster currently spends time on near-data preprocessing versus forward-pass compute. The Mathematical Deposition is the entry to the conversation; the per-buyer offload economics are a measurement, not a marketing number.
CRI-ONE commitments under the Mathematical Deposition
- Defend the priority date and claims of the substrate filing in any inter partes proceeding or third-party challenge for the life of the patent.
- Prosecute the substrate application through to issuance at CRI-ONE’s expense.
- File PCT national-phase entries in licensee-requested jurisdictions (licensee pays national-phase fees) within the 30-month PCT window.
- Maintain the patent application through statutory renewals for the full 20-year term once granted.
- Honour the deposition contract: every claim in the bundle is bisectable to a specific file plus a specific test count, and the buyer can refute the entire stack in under a minute.
- Reserve the family-data-plane integration path: CS-51-AI buyers receive priority discussion access for the Hyper Quantum Map, AutoPhi Future, and AutoPhi Quantum Battery integration tracks.
Use & Permission stance
NO IP IS SOLD. IP RETAINED BY CRI-ONE. The Mathematical Deposition conveys read-only access to the deposition contents for evaluation, citation, and internal review. The bundle is the inventor’s published method-claim package, anchored by USPTO 19/710,460 for the substrate and positioned for a CS-51 follow-on filing within the PCT window. To upgrade to Evaluation rights (full RTL) or Full Use & Copy rights (production), execute a separate agreement.
Compliance: applications are pending exam. Performance figures are design targets reproducible via the included verify.py and benchmarks/ai_pipeline_demo.py runners. AutoPhi CS-51 is a design package, not shipping silicon.
Christopher Gabriel Brown — 1341 Wellington Cove, Lawrenceville, GA 30043-5255, USA — Email: crioneaka@outlook.com. Email and postal mail only — no phone calls please.
