SCRAMJet: An Ultra Low Latency JetClass-Scale Transformer for the HL-LHC Level-1 Trigger
Abstract
Level-1 Triggers at the LHC classify collisions within a few microseconds on FPGAs, where a network is a fully unrolled circuit whose cost is area rather than time. We present SCRAMJet, a transformer encoder that replaces multi-head attention with a low-rank projection, trains under per-parameter learned fixed-point quantization, and compiles to Verilog with distributed arithmetic optimization. On the ten-class JetClass dataset it reaches 71.4 ± 0.6% accuracy with 0.157M multiply-accumulates, 6.1 accuracy points above the strongest published resource-constrained model at an eighth of the operations, and meets timing on a Virtex UltraScale+ VU13P at 147 ns using 25.7% of the logic. We also show that resource-aware training silently removes attention channels: one whose learned bit-width reaches zero still carries signal during training but is absent from the compiled circuit, so the loss never sees it go. We call these virtual channels, and a floor on the attention datalane that prevents them is worth 3.1 accuracy points at matched budget. Code and trained checkpoints are available at https://github.com/Aaddy-1/hgq-jetformer.
Related
- Low-Latency Beam-Background Suppression for the BESIII Trigger System Using Machine Learning Trained on Experimental Data
- Alkaid: A Compiler Framework for Ultra-Low-Latency Kernels on Hardware
- Low-latency axial depth estimation of 2.82 ms using light field microscopy
- Realization of movable wireless TDM-PONs via low-latency dual-stage predictive beam tracking
- Software compensation of trigger-synchronous control-frame errors in qubits and qudits
BibTeX
@misc{sharma2026scramjet,
title = {{SCRAMJet: An Ultra Low Latency JetClass-Scale Transformer for the HL-LHC Level-1 Trigger}},
author = {Aadeesh Sharma and Chang Sun and Lauri Laatu and Arianna Cox and Benedikt Maier and Alexander Tapper and Maria Spiropulu and Wayne Luk and Zhiqiang Que},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.00086},
eprint = {2611.00086},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.00086},
}
RIS
TY - UNPB
TI - SCRAMJet: An Ultra Low Latency JetClass-Scale Transformer for the HL-LHC Level-1 Trigger
AU - Aadeesh Sharma
AU - Chang Sun
AU - Lauri Laatu
AU - Arianna Cox
AU - Benedikt Maier
AU - Alexander Tapper
AU - Maria Spiropulu
AU - Wayne Luk
AU - Zhiqiang Que
PY - 2026
DA - 2026/11/01
PB - arXiv
DO - 10.48550/arxiv.2611.00086
UR - https://arxiv.org/abs/2611.00086
AB - Level-1 Triggers at the LHC classify collisions within a few microseconds on FPGAs, where a network is a fully unrolled circuit whose cost is area rather than time. We present SCRAMJet, a transformer encoder that replaces multi-head attention with a low-rank projection, trains under per-parameter learned fixed-point quantization, and compiles to Verilog with distributed arithmetic optimization. On the ten-class JetClass dataset it reaches 71.4 ± 0.6% accuracy with 0.157M multiply-accumulates, 6.1 accuracy points above the strongest published resource-constrained model at an eighth of the operations, and meets timing on a Virtex UltraScale+ VU13P at 147 ns using 25.7% of the logic. We also show that resource-aware training silently removes attention channels: one whose learned bit-width reaches zero still carries signal during training but is absent from the compiled circuit, so the loss never sees it go. We call these virtual channels, and a floor on the attention datalane that prevents them is worth 3.1 accuracy points at matched budget. Code and trained checkpoints are available at https://github.com/Aaddy-1/hgq-jetformer.
ER -