In the noisy, high-stakes world of nuclear physics, detectors are more than sensors. They’re tremulous listeners that emit streams of tiny signals when atoms rearrange, ions crash, or photons whisper from a gamma-ray shower. The critical trick is not just catching a single flash accurately, but handling thousands of signals in parallel, every nanosecond counting. It’s a choreography where speed, precision, and reliability must all keep time with the universe’s most demanding experiments.
Enter NUMEXO2, a 16-channel, 14-bit/200-MHz digitizer and processing board born at GANIL, the Grand Accélérateur National d’Ions Lourds in Caen, France. This device started life as a successor to older gamma-ray spectrometers for the EXOGAM array, but it grew into something far broader: a versatile digital brain that can live inside a detector system and perform real-time analysis, timing, and data shaping. The project is led by Charles Houarner and a collaborative team at GANIL (with partners in Spain and Italy among its authors). What makes NUMEXO2 special isn’t a single clever trick; it’s a philosophy: push the data flow forward from the moment signals arrive, without pausing to stash samples, and let a bank of programmable logic figures out the physics as the data streams along.
What NUMEXO2 is and why it exists
The NUMEXO2 module is built around a sturdy idea: digitize at the detector, then let the FPGA fabric do the heavy lifting in real time. It’s a 16-channel system arranged as one main board and four daughter boards. The heart of the module contains large FPGAs—the VIRTEX6 and VIRTEX5 families from Xilinx—that work in tandem with a Linux-running processor to produce, package, and transport the results. Each ADC channel is independent, with its own timing and dead time, so everything can run in parallel. The goal is to extract energy and time information directly from the raw waveforms and to deliver a compact, machine-readable frame that can travel to a computer cluster or a data logger without bottlenecks.
Originally designed to replace older analog VXI digitizers for EXOGAM detectors, NUMEXO2 quickly proved its flexibility. It has been adapted to a dazzling variety of detectors: HPGe, silicon strip detectors, ionization chambers, scintillators, drift chambers, and more. GANIL has produced more than 120 NUMEXO2 modules to meet the institute’s needs, and the board’s architecture was intentionally engineered to be repurposed as a general-purpose digitizer. The architectural promise is simple: a modular, high-throughput platform that can deliver high energy resolution, precise timing, and pulse-shape analysis across different detector technologies.
One of the striking design choices is the reliance on reprogrammable components, especially the FPGAs, to tailor the digital signal processing (DSP) pipelines to each detector type. The system reads out 112 LVDS data lines at a combined rate of about 44.8 Gbit/s, deserializes them, and feeds them into the DSP blocks without bogging down in memory-heavy sampling: the “samples-as-you-go” philosophy. The data frame each channel produces—energy, arrival time, and channel metadata—travels out via Gigabit Ethernet or optical links, all synchronized by a global clock and time-stamp service known as GTS, or Global Trigger and Synchronization.
A digital orchestra in a single module
The data path inside NUMEXO2 is a carefully choreographed sequence. Signals from detectors arrive as fast, delicate analog pulses. They’re digitized by four daughter boards, each holding two 14-bit, 250-MHz ADC channels, downsampled to a 200-MHz rate to align with the GTS cadence. The first FPGA (VIRTEX6) performs real-time digital signal processing, filtering, and feature extraction. It computes the energy and time information that physicists rely on, then embeds these parameters into a small, highly structured frame. A PPC440 processor running Linux on the motherboard then packages the data for transmission to the second FPGA (VIRTEX5), which attaches a universal Time Stamp (TS) and forwards the results to the data network. In other words, the board acts like a single, noisy detector’s brain, then immediately hands the result to the outside world with a precise clock and a unique time tag for every trigger event.
The heart of the processing chain is not a single algorithm but a constellation of digital signal processing blocks. On the trigger side, NUMEXO2 employs a differentiator-like step to generate an initial trigger from the raw samples, and it can optionally apply a digital Constant Fraction Discriminator (dCFD) to reduce timing walk. That is, the system can adjust when it calls a “signal” the moment it crosses a fixed fraction of its amplitude, which improves timing accuracy for pulses of different shapes. To detect low-energy events without drowning in noise, the board uses a cascade of three low-pass IIR filters with slightly offset cutoffs. These subtle filters tame high-frequency jitter without destroying the precious slow signals from large detectors, enabling reliable triggers down to tens of keV in a wide energy range.
On the energy side, NUMEXO2 uses a trapezoidal digital filter—chosen for its balance of performance and implementability in FPGA fabric. The trapezoid is tuned with parameters that shape the signal like a sculptor’s chisel: the slope k controls high-frequency rejection, and the flat-top m sets how long the energy estimate remains stable in the face of pileup. The filter is designed to be fast, recursive, and memory-efficient, so the energy calculation can run in real time for all 16 channels. Importantly, the system keeps two parallel processing streams per channel: one for the trigger (fast, low dead time) and one for energy (which may be valid or invalid depending on pileup and timing). This separation is what lets NUMEXO2 scale to high counting rates without sacrificing precision.
Time measurement is where NUMEXO2 shines. The project combines two complementary techniques to achieve sub-nanosecond timing: a linear interpolation of the dCFD zero-crossing for Tstart, and an oversampled STOP signal for Tstop. The STOP signal is sampled using an ingenious four-clock, eight-phase approach that yields an effective sampling rate of 3.2 GHz. The result is a ToF (time-of-flight) resolution in the sub-nanosecond world, with a demonstrated intrinsic time resolution around 500 picoseconds FWHM in practical tests. That combination—precise, deterministic timing with high throughput—helps researchers correlate events across multiple detectors with exquisite temporal fidelity.
All of this happens while the system preserves a clean data frame format, known as MultiFrame Metaformat (MFM). Each channel’s data frame carries not just energy and time, but the channel identifier and metadata about the detector and the trigger context. The DSP IPs are written in a mix of VHDL and Verilog, arranged in a hierarchical structure that scales gracefully from 16 channels to larger detector assemblies. The data transfer between the FPGAs uses a robust, asynchronous protocol inspired by VME, with careful handshaking to prevent data loss even when the system is lashing at full speed.
Why this matters beyond physics
If you squint at the surface, NUMEXO2 might look like a specialized piece of laboratory equipment. But its impact reaches much further. First, it demonstrates a powerful design philosophy: a digitizer that doesn’t stash samples for later processing but instead streams the essential physics in real time. That approach minimizes memory use, reduces dead time, and leverages the parallel logic resources inside modern FPGAs to extract multiple physics quantities simultaneously across many channels. In a field where a single experiment can involve thousands of detector channels, that capability is a practical game changer.
Second, the practical consequences ripple through the way experiments are run. The team at GANIL notes that NUMEXO2 dashboards an impressive counting-rate envelope—up to 50 kHz per channel, with dead times governed mainly by the trapezoidal filter parameters rather than by the digitizer itself. In high-rate environments like gamma-ray spectroscopy, the ability to maintain high efficiency at low energies (few tens of keV) while preserving energy resolution at MeV scales is a big deal. The paper’s demonstration of 2.29 keV energy resolution at 1.3 MeV, coupled with sub-nanosecond timing, shows that digital DSP can rival, and in some regimes surpass, traditional analog chains for both energy and time measurements.
Beyond the lab, NUMEXO2 hints at a future where digital electronics underpin increasingly automated, cross-detector experiments. The hardware’s architecture—modular, reprogrammable, and compatible with a broad ecosystem of detectors—suggests it could be repurposed for other domains that demand precise timing and high-throughput data acquisition. The authors even hint at upgrades, such as a newer system to replace the GTS with a more modern connectivity framework and ongoing improvements to memory and ADC performance. In short, NUMEXO2 is not just a tool for a single experiment; it’s a blueprint for how to build flexible, fast, and tightly synchronized measurement systems in physics—and perhaps in other fields that rely on rapid, multichannel data streams.
Finally, there’s a human-scale moment in this story. The NUMEXO2 project is a reminder that the most transformative hardware often starts as a clever reimagining of what a detector can do. Instead of letting data pile up in slow, bottlenecked pipelines, the NUMEXO2 approach pushes processing to the edge, right at the detector interface. It’s the digital age’s version of having a seasoned conductor standing over a chorus of musicians, ensuring every instrument contributes in perfect time. The result is not just numbers on a screen; it’s clearer, faster access to the physics those detectors were built to explore, and the potential to ask better questions, sooner.
In practice, GANIL’s NUMEXO2 has become a workhorse for a broad family of experiments, from high-resolution gamma spectroscopy to fast timing in gaseous detectors and drift chambers. A project that started with a specific spectrometer has morphed into a versatile digitizer that can be tuned to many detectors, limited mainly by the ADC clock and the software that drives the DSP blocks. The team’s careful attention to data integrity, timing alignment, and deterministic latency means scientists can trust that what they see in the data is a faithful reflection of the physical processes they’re studying—and not an artifact of the data pipeline. It’s the kind of behind-the-scenes engineering that doesn’t grab headlines, but that makes the headlines possible.
As the authors acknowledge, there are limits. Some components have become obsolete, and the 200-MHz sampling rate doesn’t match the speed of the very fastest light- or gas-based detectors. The field is already thinking about next generations—systems that push sampling rates higher, tighten timing further, and simplify integration with modern trigger and data networks. Still, NUMEXO2 stands as a vivid example of how modern physics increasingly depends on digital, real-time processing—an advance that makes experiments more agile, more coherent, and capable of weaving together many detectors into one harmonized picture of the world at its smallest scales.
At GANIL, this board isn’t just an incremental upgrade; it’s a statement about the way experimental physics can bend technology to its needs. It’s a demonstration of how a well-crafted digital instrument can bend to many detectors, how a shared clock and a common time stamp can turn scattered signals into a unified event, and how tens of billions of little measurements, stitched together in real time, can reveal the patterns that matter in the quantum dance of nuclei. In that sense, NUMEXO2 feels less like a gadget and more like a cultural shift—toward faster, smarter, and more collaborative science that learns to listen as it measures.
