DGNNFlow: A Streaming Dataflow Architecture for Real-Time Edge-based Dynamic GNN Inference in HL-LHC Trigger Systems
New hardware design slashes power consumption by 78% while processing 10x more collision data in real-time.
A research team led by Davendra Maharaj, Tu Pham, and others has unveiled DGNNFlow, a specialized hardware architecture designed to solve a critical bottleneck in next-generation particle physics. The High Luminosity Large Hadron Collider (HL-LHC) will produce 10x more collision data, overwhelming traditional offline computing and storage. Its trigger system must filter this data in real-time using complex Dynamic Graph Neural Networks (GNNs), where the relationships between particles (edges) change during computation. Existing accelerators fail because they assume static graphs. DGNNFlow introduces three key innovations: hardware support for dynamic edge embedding computation, resolution of complex data dependencies in edge-based models, and on-the-fly graph construction.
The team implemented DGNNFlow on an AMD Alveo U50 FPGA, running at 200 MHz, to benchmark against standard hardware. The results are striking: it delivered a 1.6x to 6.3x speedup over a high-end NVIDIA RTX A6000 GPU and was 3.2x to 5.1x faster than an Intel Xeon Gold CPU. More critically for edge deployment, it consumed only 22% of the GPU's power and 25% of the CPU's. This combination of high speed and ultra-low power makes it feasible to deploy sophisticated AI models directly on the detector's trigger systems, enabling real-time analysis of unprecedented data volumes. The complete implementation is publicly available on GitHub, opening the door for adaptation in other real-time, dynamic graph applications beyond physics.
- Achieves 1.6x-6.3x speedup over an NVIDIA RTX A6000 GPU while using only 22% of the power, enabling real-time edge AI.
- Solves the dynamic graph problem for GNNs with novel hardware support for changing edge embeddings, crucial for modeling particle interactions.
- Enables the HL-LHC to process 10x more collision data by moving ultra-low-latency GNN inference to power-efficient FPGA-based trigger systems.
Why It Matters
This breakthrough makes real-time AI analysis of massive scientific datasets feasible, with implications for edge computing, autonomous systems, and financial modeling.