Comparison of concepts

Since the status of the work on the three different concepts is quite different, it is hard to compare them directly, and a detailed evaluation is not possible. Nevertheless, this section compiles important information from the previous sections in Table 4.1 and compare some aspects of the three concepts next. The first point to compare is the division of the system that includes two el- ements: the number of sectors in which the detector is divided and the time- multiplexing factor. The product of these two values indicates the number of equal sector processors on which the CMS track trigger runs. Both the TMTT and the tracklet approach divide the detector only in sectors in the rϕ-plane. The TMTT approach uses the largest possible sector size limited by the detector readout scheme. In contrast, the tracklet approach uses the smallest possible sector size that is given by the curvature of 2 GeV/c particles. Both approaches divide the sector further within the sector processor to reduce the size of the data set that needs to be processed by the individual processing blocks. The processing of the smaller data sets takes less time, and the latency is minimized by parallelizing the processing. The AM approach is the only approach that splits the detector into sectors also along η. Therefore, three different types

Table 4.1.: Properties of the different CMS track trigger concepts. Status of May 2016. Sources: [88, 93, 98–103, 106]

Property AM Approach TMTT Approach Tracklet Approach

Hardware Sectors 8 × 6 = 48 8 18

Time-Multiplexing Factor 20 36 6

Sector Processors 960 288 168

Data Sharing btwn Sectors no no yes

Track Finding Method AM + 3D Projection Hough transform +

rz-Filter 3D ProjectionTracklet + Track Fitting Method PCA-based

Linearized Fit Linearized χ

2 _{Fit /}

Kalman Filter Linearized χ

2_Fit

Processing Devices FPGA and

AM ASIC FPGA FPGA

Hardware Architecture ATCA MTCA ATCA

Crates 48 120 18

Processing Board Format Double-wide FMC MTCA Board ATCA Board

of sectors (barrel, hybrid, endcap) exists in the AM approach, i.e. three differ- ent sector processors have to be developed, which may differ slightly in their hardware architecture.

The sector size has a direct effect on the data that has to be transferred to a sector processor for a single collision event: smaller sectors generate less data, but more data need to be transferred to multiple sector processors. To keep the input bandwidth small despite large sectors, the time-multiplexing factor may be increased. This can be seen clearly at the TMTT approach that has small sectors but a high time-multiplexing factor. Another way to reduce the input bandwidth is to share the sector overlap data directly between the sectors. The AM and TMTT approaches duplicate the data of the overlap regions at the DTCs. Therefore, all stubs that belong to an overlap region are duplicated. In contrast, the tracklet approach does not exchange data until the projection stage where only data from tracks crossing the boundaries have to be transferred. Another advantage of data sharing between the sector processors is that the data may be transferred over the backplane of a crate. The disadvantage of the data sharing between sector processors is an increase in latency that is caused by the additional communication step.

Although the algorithms for the data processing stages are completely different among the three concepts, two basic parts—track finding and track fitting— exist within all of them. Therefore, some blocks of the data processing chains could be adapted and integrated into any of the three concepts. Beside the track fitting block that is basically exchangeable between all three approaches, also elements of the seed generation and 3D projection can be exchanged between the AM and tracklet approaches.

The hardware implementation of the three CMS track trigger concepts differs widely. Due to the challenging requirements of the CMS track trigger, all three concepts use FPGAs for the data processing and boards with extreme data transmission capabilities. Therefore, the ATCA standard seems to be a good choice, as there is more space for optical connectors and full-mesh backplanes are available for the communication within the crate. Nevertheless, the TMTT approach builds upon an MTCA system whose boards are much smaller than the ATCA boards.

The smaller size of the MTCA boards may also be a reason for the high number of processing boards for the TMTT approach. Another reason is the conserva- tive assumption that five boards are necessary for one sector processor. One board is foreseen for these processing stages that are not implemented yet, the rz-filter and the track fitter. With the other two approaches in mind, one board

4.3. Proposed CMS track trigger concepts

will most likely be enough to execute both of them. Furthermore, one instead of two boards for the Hough transform may be sufficient. With the increasing number of logic elements in the FPGAs, the number of boards may also be reduced. The number of boards for the tracklet approach is so low because the number given is the extrapolation to the final system, i.e. the technical progress is included in the calculation.

5. A CMS Track Trigger Concept