🔬 physics

DarkFlow: Hierarchical Digital SiPM Architecture with Low-Loss Dataflow Readout for Dark Matter Detection

This paper proposes DarkFlow, a hierarchical digital SiPM architecture utilizing local data aggregation and occupancy-aware buffering to achieve ultra-low packet loss and high temporal resolution for large-scale dark matter detection, while maintaining minimal area and power overhead in a 22nm process.

Original authors: Zirui Wang, Aras Repond, Shawn Westerdale, Wantong Li

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Zirui Wang, Aras Repond, Shawn Westerdale, Wantong Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a whisper in a crowded stadium during a rock concert. That is the challenge scientists face when hunting for Dark Matter. They use massive detectors filled with liquid argon to catch the faintest "flashes" of light (photons) that might happen if a dark matter particle bumps into an atom.

To catch these flashes, they use millions of tiny light sensors called SiPMs (Silicon Photomultipliers). The problem? When a dark matter particle hits, it can trigger a massive explosion of light (a "burst") that overwhelms the sensors. Traditional systems are like a single-lane road trying to handle a hurricane of traffic: they get jammed, and most of the data gets lost.

The authors of this paper, Zirui Wang and his team, built a new system called DarkFlow to solve this traffic jam. Here is how it works, using simple analogies:

1. The Problem: The "Whisper vs. Roar" Conflict

In these detectors, there are two types of signals:

The Whisper (S1): A tiny, fast flash of light that happens in a nanosecond. You need to catch this perfectly to know when it happened.
The Roar (S2): A massive, slow burst of light that happens later. This is like a firehose of data.

Old systems tried to record every single photon individually. When the "Roar" happened, the system got so flooded with data that it couldn't keep up, losing over 80% of the information. It's like trying to write down every single word spoken by a crowd of 10,000 people at once; you'd go crazy and miss everything.

2. The Solution: A Three-Level "Smart Assembly Line"

DarkFlow organizes the sensors into a hierarchy, like a company with a clear chain of command, to handle the data efficiently.

Level 1 (The Neighborhood Watch): Instead of reporting every single photon, a small group of 16 sensors (a "tile") acts as a neighborhood watch. They count how many photons hit them and group them into a simple code (like "Low," "Medium," or "High" traffic). They don't send a report for every single car; they just send a summary of the traffic flow.
Level 2 (The City Block): These tiles pass their summaries to a "City Block" manager. This manager adds a timestamp. But here's the trick: instead of writing down the exact time for every single event (which takes up too much space), it writes down the difference in time from the last event. It's like saying, "The next car arrived 2 seconds after the last one," rather than writing "12:00:01, 12:00:03, 12:00:05..."
Level 3 (The Highway Hub): All the City Blocks send their data to a central highway. This hub uses a clever "backpressure" system. If the exit road (the computer reading the data) gets too busy, the highway doesn't just crash. It sends a signal back up the line saying, "Slow down!" The lower levels pause sending new data, but they don't lose what they already have. They hold it in a temporary waiting room (a buffer) until the road clears.

3. The "Smart Waiting Room" (eDRAM Buffer)

When the "Roar" of light hits, the system needs a massive waiting room to hold all the data before it can be sent out. DarkFlow uses a special type of memory called eDRAM.

Usually, memory needs to be "refreshed" (wiped and rewritten) constantly to keep the data from fading, even if there's nothing important in that spot. This is like a librarian constantly re-shelving empty books just to make sure the shelves are clean. It wastes time and energy.

DarkFlow uses a "Smart Librarian" (Occupancy-Aware Refresh). This librarian only checks the shelves where books (data) actually exist. If a shelf is empty, they ignore it. This makes the system 2.14 times more efficient at keeping data safe during a crisis.

4. The Results: No Data Lost

The team tested this design using a computer simulation and built a prototype chip.

The Traffic Test: When they simulated a massive burst of light (billions of photons per second), old systems lost more than 80% of the data. DarkFlow lost almost nothing.
The Size and Power: The entire digital brain of this system is tiny. It takes up less than 1% of the total space of the detector and uses very little power. This is crucial because the detectors are submerged in super-cold liquid argon; if the electronics get too hot, they would boil the liquid and ruin the experiment.

Summary

DarkFlow is a smart, hierarchical traffic control system for light sensors. Instead of trying to record every single photon individually (which causes a traffic jam), it groups them, uses smart time-stamping, and has a "brake system" that prevents data loss when the traffic gets too heavy. It allows scientists to listen for the faintest whispers of dark matter, even when the stadium is roaring.

Technical Summary: DarkFlow – Hierarchical Digital SiPM Architecture with Low-Loss Dataflow Readout for Dark Matter Detection

Problem Statement
Direct dark matter detection experiments, particularly those utilizing dual-phase liquid argon time projection chambers (LAr-TPCs), require large-scale photon sensing arrays with hundreds of thousands of synchronized readout channels. These detectors must capture faint scintillation photons from rare Weakly Interacting Massive Particle (WIMP) collisions. While Silicon Photomultipliers (SiPMs) offer superior radiopurity and high integration density compared to traditional photomultiplier tubes, their large-scale integration presents a critical architectural conflict. The system must simultaneously resolve two distinct signal modalities: the prompt, nanosecond-level S1 scintillation pulses requiring fine timing resolution, and the delayed, microsecond-level S2 electroluminescence pulses which generate high-flux bursts exceeding 10 Gbps. Existing digital SiPM readout architectures struggle to preserve nanosecond temporal resolution for sparse events while maintaining data integrity during high-intensity photon bursts. Furthermore, the readout system operates under extreme physical constraints, including a strict milliwatt-level power budget per channel to prevent liquid argon boiling and prohibitive limits on uplink bandwidth due to thermal and cabling constraints.

Methodology
The authors propose DarkFlow, a hierarchical digital SiPM architecture designed within a unified dataflow framework to address the bandwidth and power constraints of large-scale detectors. The architecture is structured into three levels:

L1 (Sensing Unit): A 4×4 SPAD array where analog signals are summed and quantized using three comparators with distinct voltage thresholds ( $V_{th1} < V_{th2} < V_{th3}$ ). This generates a 3-bit thermometer code mapped to a compact 2-bit digital energy level. Crucially, the lowest threshold ( $P_{th}$ ) is programmable, allowing for local zero-suppression (e.g., setting $P_{th}=2$ to suppress dark counts) or dynamic elevation during massive S2 bursts to prioritize high-intensity signals.
L2 (Pixel Tile): Integrates 16 L1 modules. It captures the arrival time of the first photon with 2-ns resolution and subsequently aggregates spatial-energy data at a coarser 10-ns frame rate. L2 nodes employ a systolic readout mechanism, daisy-chaining data to the end of the row with single-slot skid buffers to avoid arbitration collisions.
L3 (Row-Level Aggregation): Utilizes a dual-bank systolic array to manage even and odd row groups independently. It aggregates data from L2 nodes and manages a 128-kB burst-absorbing FIFO based on embedded DRAM (eDRAM).

Key Innovations

Hierarchical Packet Design: DarkFlow decouples absolute timing from high-frequency local events. Instead of appending an absolute timestamp to every photon, the system uses a compact 16-bit packet format containing a 15-bit relative time offset (2-ns resolution) and a 7-bit sequence index (10-ns steps). A separate 32-bit global time packet is broadcast periodically to anchor these relative timestamps, reducing per-event bandwidth overhead by an order of magnitude.
Consumer-Driven Backpressure: To prevent data loss during congestion, the system implements a layered backpressure mechanism. When downstream buffers fill, stall signals propagate upstream from L3 to L2 and finally to L1, pausing data generation only when local FIFOs are saturated. Single-slot skid buffers hold in-flight packets during these stalls, ensuring no data is dropped.
Occupancy-Aware Refresh: Addressing the limited data retention time of eDRAM during high-flux bursts where refresh cycles are scarce, the authors propose a refresh mechanism that confines the refresh pointer within the active data window (bounded by read and write pointers). This ensures every refresh cycle targets valid data, avoiding the waste of cycles on empty addresses inherent in global refresh schemes.

Results
The architecture was evaluated using a custom Python-based simulator and hardware implementation in GlobalFoundries 22nm technology at 500 MHz.

Data Loss Performance: Under simulated S2 burst conditions of $10^9$ to $2 \times 10^9$ hits/s, conventional serial and Address Event Representation (AER) architectures exhibited packet loss exceeding 60% and up to 80%. In contrast, DarkFlow maintained ultra-low packet loss. With a photon threshold of $P_{th}=1$ , it remained lossless up to $2 \times 10^9$ hits/s; with $P_{th}=2$ , the lossless range extended beyond $10^{10}$ hits/s.
Refresh Efficiency: The occupancy-aware refresh mechanism achieved a 100% refresh hit rate (targeting only valid data) compared to 46.8% for conventional global refresh. This resulted in a 2.14× improvement in effective refresh rate, significantly increasing the data retention margin.
Hardware Overhead: The complete DarkFlow digital readout datapath occupies 0.255 mm² and consumes 78.7 mW. When compared to a typical 256×128 SPAD array (approx. 29.5 mm²), the digital logic accounts for less than 0.86% of the total detector area. The power dissipation density is approximately 2.6 mW/mm², which complies with the thermal budget for immersed electronics in liquid argon environments.

Significance
The paper claims that DarkFlow successfully resolves the fundamental conflict between the need for nanosecond-level timing resolution and the requirement to sustain high-throughput data integrity during billion-photon events. By unifying sparse S1 and burst-heavy S2 traffic patterns into a single deterministic dataflow framework, the architecture enables the scaling of SiPM arrays to hundreds of thousands of channels without violating strict power and bandwidth constraints. The work demonstrates that hierarchical aggregation, compact relative-time encoding, and intelligent buffer management can preserve spatial and temporal information in extreme photon flux conditions where existing architectures fail, thereby facilitating more sensitive direct dark matter detection experiments.

1. The Problem: The "Whisper vs. Roar" Conflict

2. The Solution: A Three-Level "Smart Assembly Line"

3. The "Smart Waiting Room" (eDRAM Buffer)

4. The Results: No Data Lost

Summary

More like this