Aien Research · Preprint, not peer reviewed
Reverse Engineering the CUDA Software Stack: An Independently Implemented GPU Machine-Code Toolchain for the NVIDIA GB10
M. Drake Stapleton · Version 2, 26 October 2026 · DOI 10.5281/zenodo.23677465
Read it
- paper.pdf (sha256
3f52d023c7b36e4d6a3edcc8d4bf8d39bcac6b549c38fd9ec704940bb424d392) - paper.tex, the source (sha256
12bdd38d1b6e83c7c929527d7b748cbbbe4fbcb74bf21cac932c9944817f7af2) - SHA256SUMS
- The record on Zenodo: doi.org/10.5281/zenodo.23677465. These files are byte-identical to the ones uploaded there. License: CC BY 4.0.
Abstract
"I describe CUDA++, a research-stage GPU machine-code toolchain and runtime for the NVIDIA GB10 in a DGX Spark. It provides an execution path for demonstrated workloads without NVIDIA's proprietary CUDA user-space compiler, runtime, driver library or mathematics libraries. It is not a drop-in implementation of the CUDA language, APIs or library ecosystem. The path still uses NVIDIA's open Linux kernel modules and signed GPU firmware."
Versions
This is version 2, the current one. Version 1 (22 October 2026, DOI 10.5281/zenodo.23578453) is superseded: version 2 narrows its claims and lists every correction. Cite version 2. The concept DOI 10.5281/zenodo.23578452 always resolves to the latest version.
The code
The software the paper describes is CUDApp, published on Radicle from my own node: rad clone rad:z4RxD8vXigmWg61nM8miQ8DVPAXEP. What I learned along the way, finding by finding, is in the DGX Spark field manual. All my other papers are on the research page.
Cite
Stapleton, M. Drake (2026). Reverse Engineering the CUDA Software Stack:
An Independently Implemented GPU Machine-Code Toolchain for the NVIDIA GB10.
Version 2. Zenodo. https://doi.org/10.5281/zenodo.23677465