Fast inference of deep neural networks in FPGAs for particle physics 论文

详细信息

发表期刊/会议

Journal of Instrumentation

发表日期

2018-07-27

发表年份

2018

摘要

processing of collision events. The task of the real-time processing is to filter events to reduce data rates to manageable levels for offline processing is called triggering. It is typically performed in multiple stages Because of the extreme input data rates and size of the data buffers, the first stage, Level-1 (L1), of data processing typically uses custom hardware with ASICs or, increasingly, FPGAs, to handle the initial data rate using pipelined algorithms with latencies of hundreds of nanoseconds totalling microseconds. The second stage of triggering, High Level Trigger (HLT), uses commercial CPUs to process the filtered data in software with longer latencies on the timescale of hundreds of milliseconds in total.