Nanosecond machine learning regression with deep boosted decision trees in FPGA for high energy physics

Abstract

We present a novel application of the machine learning / artificial intelligence method called boosted decision trees to estimate physical quantities on field programmable gate arrays (FPGA). The software package fwXmachina features a new architecture called parallel decision paths that allows for deep decision trees with arbitrary number of input variables. It also features a new optimization scheme to use different numbers of bits for each input variable, which produces optimal physics results and ultraefficient FPGA resource utilization. Problems in high energy physics of proton collisions at the Large Hadron Collider (LHC) are considered. Estimation of missing transverse momentum (ET miss) at the first level trigger system at the High Luminosity LHC (HL-LHC) experiments, with a simplified detector modeled by Delphes, is used to benchmark and characterize the firmware performance. The firmware implementation with a maximum depth of up to 10 using eight input variables of 16-bit precision gives a latency value of 𝒪(10) ns, independent of the clock speed, and 𝒪(0.1)% of the available FPGA resources without using digital signal processors.

Machine learning evaluation in the Global Event Processor FPGA for the ATLAS trigger upgrade

“…The software package fwXmachina is used for implementing boosted decision tree-based machine learning algorithms onto FPGAs for high-energy physics applications [15][16][17]. Similar to hls4ml, it uses Vivado HLS to convert the model into RTL.…”

“…The implementation in the fwX regression studies was originally performed using public Delphes samples [27] described in ref. [16].…”

“…The eight input variables are described in ref. [16]. The regression model is configured with 40 trees, a tree depth of 6.…”

Ultrafast jet classification at the HL-LHC

“…However, these tools aim at implementations that are not optimized for the L1T systems, and they do not necessarily support the neural network architectures studied here. Conifer [29] and fwXmachina [30][31][32] feature custom implementations of boosted decision trees on FPGAs, which achieves the desired L1T constraints, but cannot be extended to neural networks. LL-GNN [33] proposes a domain-specific low latency hardware architecture for processing GNNs in high energy physics, which involves many manual optimizations.…”

Trees versus Neural Networks for enhancing tau lepton real-time selection in proton-proton collisions

“…Events passing the L1 trigger are further filtered by the high-level trigger (HLT), based on CPUs and GPUs, where more advanced algorithms can be used. However, there have been recent developments in engineering that allow for the implementation of machine learning (ML) methods on FPGAs within the timing and resource constraints necessary at L1 [6][7][8] . These advancements represent a paradigm shift, empowering L1 with the capability to execute advanced algorithms previously exclusive to the HLT, thereby enhancing event selection efficiency and broadening the scope of physics analyses within the experiments.…”