Generic Optical Link Data Processor

The ATLAS Level-1 Calorimeter Trigger algorithms are mainly based on the analysis of localized objects (e.g. electromagnetic clusters or jets) that are identified by sliding-window algorithms. On both phase 1 and phase 2 of the trigger upgrade there will be a need for a topological post-processing of objects identified by the sliding-window processors. There are two fundamentally different approaches to processing data containing information on the event topology:

For algorithms requiring the presence of two or more objects it is sufficient to present the n most significant objects of each object type to the top level processor. Each object needs to be tagged with position, type and significance. The latter must be on a monotonous scale, such that a simple arithmetic comparison can determine the largest objects.

In hardware, this scheme maps onto a simple tree-like structure where a large number of object data is entering a processor, a sort algorithm determines the most significant ones, and a small number of objects is passed on to the final decision logic. Such architecture can easily be resolved into a large number of simple, identical stages of compare-and-select operations. The only limitation here comes from the latency induced at chip and module boundaries, in particular when high speed serial links are employed. Therefore excessive partitioning into microscopic sub-processors ought to be avoided.

For algorithms requiring the presence of one object in conjunction with the absence of another one (e.g. veto on both a jet and an electromagnetic cluster in a given region), the scheme described above is not suitable. Here, along with each object identified, we would need the full set of data describing all sections of the solid angle that might possibly be associated with this object. In the extreme, such algorithms would require all available data to be mapped onto a single processor. Some level of data reduction might be possible, once the detailed algorithms are known. However, for a generic two-dimensional ( h , j ) topological processor, data transport into a compact and powerful processor is a major challenge.

Due to the algorithms to be used, the phase-1, as well as the phase-2 top level processors will most likely require both tree-like and two-dimensional processors. It is probably worth pointing out that for phase-2 a new sliding window processor will have to be built, which is just another two-dimensional processor as described above.

Topological data will be carried on optical fibres. Since it will not be possible to route all data into a single processor chip, incoming data need to be copied to several chips, or even several modules. Data can effectively be replicated in three different ways: electrical duplication at the source, passive fibre couplers (splitters), and electrical duplication at the sink. Retransmission at the sink would rather be avoided, for reason of latency.

Dependent on the link replication scheme and on the module form factor there are various options how to route the optical links into the module. Front panel opto connectors are messy, once the number of links per module gets large. Rear transmission modules can be particularly useful, if links are to be replicated to neighbouring slots electrically. However, at very high link speeds the scope for electrical routing across backplanes and connectors is very limited. From both logistical and signal integrity point of view the use of high-density blind-mate optical backplane connectors is probably the most favourable approach.

For modules mainly based on optical connectivity, no standard form factor is known. However, slot pitch is impractically dense for VME modules with sub-modules, in particular when taking thermal issues into account. Also, 9Ux40cm module dimensions are at the very limit for a densely packed, densely routed module. Track length limitations at high speeds might force the designer to squeeze high-speed sections of the design into just a fraction of available space. ATCA, on the other hand, offers the advantage of a considerable backplane height, moderate board area, and increased slot width. Rear transition modules are integral part of the ATCA specs. Power dissipation limits on RTMs were low, but are currently being lifted. The concept of -48V primary supply is superior to the VME supply scheme. A combination of ATCA standard crate mechanics, ATCA power distribution, and a custom backplane in the remaining back space would be an attractive scheme for opto fibre based module connectivity.

It is planned to build an ATCA sized demonstrator module, based on Xilinx Virtex-6 FPGAs and high density blind mate opto connectors. Xilinx V6 road map suggests a time line with module manufacturing in early 2010. By then a module could be equipped with either engineering samples or production chips. The exact dates for availability of production silicon cannot currently be predicted. High density blind mate opto connectors are available from Molex and other companies. It would make sense to procure samples of opto connectors in the next couple of months. If the backplane optics concept fails to convince, electro-optical RTMs would probably be preferred over front panel connectivity.

Main goal of the demonstrator project is the proof of concept of extensive high speed link replication. Therefore it is planned to equip the demonstrator with a larger number of rather small FPGAs. By the time a prototype will be built, details of the algorithms will be known. Therefore one would at that stage either convert the design to fewer large devices, so to reduce link replication. Or, if the algorithms turn out to require large amounts of logic resources, one would have to go for many large devices.

The demonstrator project is a priori not dependent on any specific make of processors or opto link components. Xilinx was chosen, since tools and expertise are available in the labs. By the time of prototype development the market situation should be assessed so as to choose suitable components of known (Xilinx, Altera, Lattice) and possibly unknown manufacturers of high performance FPGAs.

Implementation details

The demonstrator will be built in ATCA form factor. The power supply zone (Z1) will be compatible to a standard ATCA crate. Zones 2 and 3 will not be ATCA compliant. As outlined above, optical fibres will be routed directly onto the module. So as to achieve the required level of connectivity, the largest part of backplane height needs to be dedicated to fibre connectors. Since all real-time backplane connectivity will be made optically, it does probably make sense to run crate control, environmental, and timing data on fibres as well. A single electrical cable connection to the CTP will be made on the front panel. Alternatively a single block of differential electrical backplane I/O could be made available, if fibre routing density in the remaining part of the backplane is sufficiently high.