EIC Streaming Data Path

The ePIC detector has no hardware trigger, so background has to be thrown away inside the readout chain, on data that is still flowing. That sets the two hard limits this benchmark scores against: a bandwidth cliff at every stage, and a 0.6 ms time-frame to finish in.

Bandwidth: where the data has to shrink

Front-endASICs · FEB / DCMO(2 Pbps)÷200RDOreadout FPGA boardO(10 Tbps)÷20DAM / EBDCaggregation · AI filter siteO(0.5 Tbps)÷5Buffer Boxswitch · Echelon-0 · STFsO(0.1 Tbps)Echelon-1offline / near-line recofiles, not a streamstreaming, no hardware trigger · must keep pace with the beam in real timenot latency-boundproposed site for the AI filter
Rates are the figures quoted under each stage on the Genesis Fast-ML slide (M. Liu, 5 Aug 2026); ÷ factors are the ratios of those order-of-magnitude figures, and arrow thickness is drawn on a log scale. Cutting 10 Tbps to 0.5 Tbps at the DAM is where an AI background filter has to earn its place; the 0.1 Tbps out of Echelon-0 matches the ~100 Gbps ePIC detector stream in J. Landgraf's Echelon 0–1 concept.

Latency: what fits inside a time frame

1 µs10 µs100 µs1 ms10 ms100 ms1 s0.6 ms time-frame budgettoo slow to run inline2 µs integration windowHEPT on CPU · 5.1 msHEPT on GPU · 2.1 msEICrecon reco, 12 ms filter in front · 278 msEICrecon reco, stock · 607 ms
Log time axis. The rule-based reference (EICrecon) takes ~607 ms per event on one CPU thread (278 ms with the AI hit filter in front) in the 15 Jul 2026 streaming-readout study (slide 31, “Tracking speedup”, 2026-07-15 Streaming Readout AI Background Suppression; preliminary, measured on the 26.07.0 DIS 10x275 minQ2=1000 campaign, not this benchmark’s April 26.04.1 10x100 sample), about a thousand times over the 0.6 ms budget. That is why it is the accuracy reference, not a deployable entry. The HEPT CPU/GPU timings are the upstream model on its own dataset, not this benchmark's.

Limits a submission must live inside

Unconfirmed: the AI filter's placement (DAM/EBDC vs. RDO) is the slide's proposal, not a decided design. The 0.6 ms budget is the proposal's working number. The EICrecon timings come from a different study configuration and have not been re-measured on this benchmark's data.