mirror of
https://github.com/lux-apps/papers.git
synced 2026-10-30 18:15:21 +00:00
zeekayandHanzo Dev
178febad68
Every rendered "T-Chain" / "ThresholdVM" / "teleportvm" reference across the paper corpus is rewritten to the correct canonical chain by concern: threshold signing / MPC / bridge custody / key recovery -> M-Chain (MVM); FHE / encrypted compute / threshold decryption -> F-Chain (FVM); teleport / cross-chain messaging -> B-Chain (bridgevm); consensus Pulsar stays Q-Chain. Section lux-trust-kernel/03-t-chain-threshold-governance.tex renamed -> 03-f-chain-encrypted-policy.tex (with \input + xref fixes). Braided "T-Chain" sections split by concern (M vs F). Only lp-020's chain-separation rationale and the shared lux/chains/thresholdvm library path retain the name, to record the removal. All changed papers rebuilt (pdflatex, rc=0, no undefined refs). ยง09 of lux-dao-governance-framework reframed: fabricated "production deployment" metrics ($50M/$240k/$12M, survey %s, "as of Oct 2024") relabeled as illustrative design configurations with no invented replacements; reality noted (DAO treasury is a 1/1 Safe); conclusion echo + one "first production" overclaim softened. This commit also captures concurrent corpus workstreams on the same tree. Co-authored-by: Hanzo Dev <[email protected]>
84 lines
2.9 KiB
TeX
84 lines
2.9 KiB
TeX
\documentclass[11pt,a4paper]{article}
|
|
\usepackage{../shared/luxcover}
|
|
\usepackage[utf8]{inputenc}
|
|
\usepackage[T1]{fontenc}
|
|
\usepackage{amsmath,amssymb,amsfonts}
|
|
\usepackage{graphicx}
|
|
\usepackage{hyperref}
|
|
\usepackage{booktabs}
|
|
\usepackage{enumitem}
|
|
\usepackage{xcolor}
|
|
\usepackage{siunitx}
|
|
\usepackage{listings}
|
|
\usepackage{url}
|
|
|
|
\hypersetup{colorlinks=true,linkcolor=black,citecolor=blue,urlcolor=blue}
|
|
|
|
% Allow looser line breaking on \texttt to reduce overfull boxes from
|
|
% long unbreakable identifiers (cuda_op_*, etc.).
|
|
\sloppy
|
|
\hbadness=10000
|
|
\hfuzz=20pt
|
|
|
|
\lstset{
|
|
basicstyle=\ttfamily\scriptsize,
|
|
breaklines=true,
|
|
frame=single,
|
|
showstringspaces=false,
|
|
}
|
|
|
|
\title{GPU-Native Verify on Grace Blackwell:\\
|
|
Measured Numbers, Honest Gaps, and the\\
|
|
Wire-Up Contract for the Consensus Hot Path}
|
|
\author{Lux Industries, Inc. \\
|
|
\texttt{[email protected]}}
|
|
\date{2026-06-03}
|
|
|
|
\begin{document}
|
|
\luxcoverpage
|
|
|
|
\begin{abstract}
|
|
LP-203 specifies a GPU-native verify path where signature
|
|
verification, lattice signing, Merkle-root computation, and
|
|
threshold aggregation run as CUDA / Metal / ROCm kernels reading
|
|
ZAP buffers directly from unified memory. Bytes arrive at the GPU
|
|
through one of three transports -- kernel sockets, DPDK with a
|
|
hugepage UMA mempool, or GPUDirect RDMA -- and there is no
|
|
\texttt{cudaMemcpy} between the wire and the kernel. This paper
|
|
expands on LP-203 (\texttt{lps/LP-203-gpu-native-verify.md}) with
|
|
the first round of \emph{measured} numbers on NVIDIA GB10 Grace
|
|
Blackwell sm\_120 (DGX Spark, 2026-06-04), the wire-up architecture
|
|
(\texttt{luxfi/accel} dispatch, UMA via \texttt{cudaMallocManaged},
|
|
DPDK PMD, GPUDirect RDMA), an honest inventory of the gaps the
|
|
matrix bench exposed (composite signature-verify wrappers for
|
|
Magnetar/Pulsar/Ed25519/sr25519 return
|
|
\texttt{LUX\_BACKEND\_ERROR\_NOT\_SUPPORTED}; the matrix bench observed mean
|
|
SM utilization of \SI{11.9}{\percent} -- the GPU was not exercised
|
|
by any path), the projected numbers for the wrappers once they
|
|
land (explicitly tagged ``projected, not measured''), and the
|
|
composition contract with LP-201 (transport selector), LP-205
|
|
(pipelined Quasar), LP-217 (cert profile modes), and LP-218 (P3Q
|
|
at slot \texttt{0x012205}). Measured BLS12-381 EIP-2537 pairing on
|
|
GB10: \SI{19}{\micro\second} per pair sustained,
|
|
\num{49197}~pairings per second at \SI{20}{\watt}, ~\num{12300}
|
|
pairings per watt. Measured keccak256: \SI{11.38}{\mega\hertz}
|
|
sustained at $N{=}4096$. Composite signature-verify is the gap to
|
|
close before the LP-203 throughput tables turn fully empirical.
|
|
\end{abstract}
|
|
|
|
\tableofcontents
|
|
\newpage
|
|
|
|
\input{sections/01-introduction}
|
|
\input{sections/02-architecture}
|
|
\input{sections/03-measured-benches}
|
|
\input{sections/04-honest-gaps}
|
|
\input{sections/05-projections}
|
|
\input{sections/06-composition}
|
|
\input{sections/07-abi-input-bounds}
|
|
|
|
\bibliographystyle{plain}
|
|
\bibliography{biblio}
|
|
|
|
\end{document}
|