UFAIRA
Ultra-Fast AI Inference for Real-Time Applications
FPGA-based AI Inference

From particle physics
to high-speed networking.
Nanosecond latency.

We work with leading research institutions and industry partners to deploy custom AI models on our FPGA hardware. Not only beating traditional approaches but outperforming all state-of-the-art technologies in latency, throughput and efficiency.

Contact Us
Prototype in Testing
Particle Physics

Single Particle Counting & Localization

GSI Helmholtzzentrum für Schwerionenforschung, Darmstadt

Accurate particle counting and precise arrival time determination are fundamental requirements in experimental nuclear and particle physics , as well as in accelerator beam diagnostics. In high-rate environments such as those encountered at the GSI Helmholtz Centre for Heavy Ion Research, scintillator-based single particle counters frequently suffer from pile-up events, where multiple particles arrive within short time intervals and produce superimposed pulse shapes at the detector output.-[M. Hamdan, T. Habermann, 2026] The core challenge: When two or more particles pass the scintillator detector in close temporal proximity, the emitted signals overlap and superimpose. Classical methods cannot reliably disentangle them. Our CNN was trained specifically to decompose superimposed signals and localize each particle, achieving a temporal resolution below 100 picoseconds — a capability never demonstrated before in real-time hardware.

A fully functional prototype is currently undergoing extended validation tests at the GSI facility.

<300 ns
Inference Latency
120 Gb/s
Data Rate
<100 ps
Timing Resolution
10 GSa/s
Sample Rate
12 bit
ADC Resolution
Convolutional Neural Network FPGA Synthesis Real-Time Signal Processing Particle Accelerator Superposition Disentanglement Sub-ns Timing
Key Achievements
World-first real-time ML-based single particle counting at 120 Gbit/s data rate
Beats all traditional methods (threshold, pulse-fit) in accuracy, resolution and throughput
Disentangles overlapping particle signals that are indistinguishable to classical signal processing
Prototype validated and in active testing at the GSI accelerator facility

References

[M. Hamdan, T. Habermann, 2026]
M. Hamdan, T. Habermann (2026). Exploring Convolutional Neural Networks training strategies for Pile up correction in single particle counting IPAC 17th International Particle Accelerator Conference, Deauville France.