ionis-cuda

CUDA signature embedding engine for the IONIS project.

Overview

Generates float4 embeddings from WSPR spot data and solar indices using CUDA kernels on NVIDIA GPUs.

Pipeline:  wspr.bronze + solar.bronze  ──▶  bulk-processor (CUDA)  ──▶  (destination table)
Hardware:  RTX PRO 6000 (96 GB VRAM) — single-pass processing
Wall time: ~45 min on Threadripper 9975WX

Not currently wired into the IONIS pipeline

bulk-processor wrote to wspr.silver, which was dropped on 2026-09-22 holding zero rows. It was probably not always empty — a QA rebuild recorded 4.43B rows on 2026-02-07 — but ClickHouse's logs only retain back to 2026-09-06, so when it emptied cannot be established. That a table could lose four billion rows unnoticed for seven months is the finding. It could, because nothing read it: this tool is not packaged in any RPM, has no systemd unit, and runs only by hand, while all fourteen gold populate scripts read wspr.bronze directly. The medallion chain the docs described — bronze → silver → gold — was a design, not the build. The build is bronze → gold.

The CUDA engine itself is sound and is kept. What it lacks is a consumer. Before running it again, decide where its output goes and what reads it; sql/01-model_features.sql still carries the original schema for reference, but creates nothing.

Current lineage for every table in the lab: ionis-core/docs/DATA-DICTIONARY.md

Components

Component	Description
bulk-processor	Main CUDA embedding generator — reads ClickHouse, writes float4 embeddings. No destination table at present — see above.
wspr-cuda-check	Quick GPU capability check utility
src/cuda/	CUDA kernels for embedding computation
src/engine/	Processing engine and batch orchestration
src/io/	ClickHouse I/O with Maidenhead grid conversion

Requirements

• NVIDIA GPU with sufficient VRAM (tested on RTX PRO 6000, 96 GB)
• CUDA 12.8+ toolkit
• NVIDIA driver 570+
• CMake 3.28+
• ClickHouse with populated wspr.bronze and solar.bronze

Building

cd build/cmake
cmake -B build -DCMAKE_CUDA_ARCHITECTURES=120
cmake --build build

# Or use the top-level Makefile
make all

Usage

# Run the bulk processor (default host: 192.168.1.90)
bulk-processor --host 192.168.1.90

# Environment variable override
CH_HOST=10.60.1.1 CH_PORT=9000 bulk-processor

Installation

From COPR (Recommended)

sudo dnf copr enable ki7mt/ionis-ai
sudo dnf install ionis-cuda

Upgrading from ki7mt-ai-lab-cuda

The ionis-cuda package includes Obsoletes: ki7mt-ai-lab-cuda for seamless upgrade:

sudo dnf copr enable ki7mt/ionis-ai
sudo dnf upgrade --refresh

From Source

git clone https://github.com/IONIS-AI/ionis-cuda.git
cd ionis-cuda
make all
sudo make install

Related Repositories

Repository	Purpose
ionis-core	DDL schemas, SQL scripts
ionis-apps	Go data ingesters (WSPR, solar, contest, RBN)
ionis-training	PyTorch model training
ionis-validate	Model validation suite (PyPI)
ionis-docs	Documentation site

License

GPL-3.0-or-later — See COPYING

Author

Greg Beam, KI7MT

Links

• COPR: https://copr.fedorainfracloud.org/coprs/ki7mt/ionis-ai/
• Issues: https://github.com/IONIS-AI/ionis-cuda/issues