After the upgrade of the Hera machine, an increase of luminosit by a factor 5 is expected. In order to meet this new requirement and to keep data input flow to the network still manageable, an improved preprocessor has been developed extracting the physically relevant informations while increasing rejection
power. We will describe the new neural post processor (DDB2) and focus especially on the algorithmics principles which generate more physics oriented quantities (jets, particles,...) from the full granularity of the L1 trigger processors. This principle relies on intelligent preprocessing of data sent to the neural net and is executed in 4 steps :
1. The clustering unit, which aim is to localize regions of interests within a subdetector in the (the,phi) plane.
2. The matching unit, taking over the correlations and associations of informations from all subdetectors in order to build physical objects defined by their topological vicinity.
3. The sorting unit, which consists in ordering different lists of objects according to specific criterions.
4. The post-processing unit, generating variables destined to be exploited by the neural network.
A general overview of the hardware implementation of such algorithms will then be discussed, notably the use of the current, fast FPGA technology to conjugate parallelism and speed and meet the timing constraint of 8µs to process all algorithms.