SparkBroker.
The GPU-Based MQTT Broker for Real-Time Industrial Sensor Data.
Standard industrial MQTT brokers are too slow to see the heartbeat of a high-speed asset — leaving you blind to the micro-ailments that lead to failure. SparkBroker breaks that ceiling.
By extending HiveMQ CE with NVIDIA CUDA processing on DGX Spark hardware, we process up to 27 million messages per second in thresholding mode and run real-time signal-analysis algorithms like sliding Fourier transform inline against streams of 5 million messages per second — on sovereign, affordable hardware.
Two Operating Modes
Configure SparkBroker for raw throughput or in-flight analysis — the same hardware, two different precision profiles.
Thresholding Mode
GPU-accelerated rules engine filters raw sensor traffic against threshold conditions, forwarding only the messages that matter to the downstream HiveMQ broker.
Algorithmic Mode
GPU-resident DSP pipelines — sliding Fourier transform and other real-time signal-analysis algorithms — run inline against the message stream for in-flight anomaly detection.
The Vision: Real-Time Anomaly Detection at Sensor Speed
SparkBroker isn't just a faster broker. It's the foundation for a class of industrial AI that operates at the resolution failures actually happen at.
Real-Time Anomaly Detection
Detect bearing degradation, valve chatter, motor harmonic shifts, or process excursions in the moment they begin — not in the next-day report.
Closing the Resolution Gap
Standard industrial brokers sample your factory's physics at the wrong rate. SparkBroker captures the heartbeat of high-speed assets at the resolution failures actually live at.
Noise Filtering at the Edge
Industrial signal channels are mostly noise. The GPU rules engine filters the noise before it ever reaches your historian, dashboards, or AI models — saving downstream cost.
Architecture
How the pieces fit together.
GPU-Resident Message Engine
Extends HiveMQ CE MQTT Broker with NVIDIA CUDA processing on DGX Spark hardware. Messages never leave GPU memory during inspection.
High-Bandwidth Ingestion
Supports RDMA and UDP over QSFP56 ports for the raw bandwidth required to keep up with high-frequency industrial sensor streams.
Sub-Millisecond Decisions
Decisions about which messages forward, drop, or trigger anomalies happen in microseconds — faster than the time between two PLC scan cycles.
HiveMQ Compatibility
Operates as a drop-in extension to HiveMQ CE. Existing MQTT clients, Sparkplug B publishers, and Unified Namespace consumers connect without changes.
See It in Action
Why This Matters
The future of industrial sensor data isn't more dashboards — it's real-time anomaly detection at sensor speed. Pumps, valves, motors, and reactors fail in patterns that show up in the signal milliseconds before they show up in the process. By the time the historian writes the row, the variance is already operator-visible.
SparkBroker is built to catch those patterns the moment they emerge — not in the next-shift report. We see this as the substrate that AI-driven predictive maintenance and process-anomaly systems will run on top of for the next decade. We're building it now.
Interested in Early Access?
SparkBroker is in active development. We're talking with manufacturers, integrators, and research groups about pilot deployments and joint development. If your use case lives in the gap between sensor speed and historian latency, we'd like to hear from you.