In the 20th century high end correlators were built with ASICs (VLA, Westerbork, Australia Telescope). As we entered the 21st century FPGAs came to the fore (eVLA, CABB, etc.) but due to the huge data flows they needed massive dedicated data transport systems. The CASPER group followed a different path and used network switches for data transport but this imposed a significant extra cost and for large systems such as MeerKAT with considerable effort and interaction with the switch vendor to get the system to work. The last 5 years has seen the appearance of a new generation of Ethernet switches (In-Network Processors) where the data plane is fully programmable using languages such as P4 and OpenFlow. These languages not only allow precise routing of packets based on metadata within the Ethernet packets but also open the door to a precise monitoring of the data flow. Thus, the designer now has the control available with older designs but without the complexity and with more advanced features.

The second revolution was the combination of FPGAs and High Bandwidth (attached) Memory (HBM). HBM memory has an I/0 bandwidth that is more than 30 times that of a 100GbE link. This allows the full data to be buffered to memory multiple times. Each stage of processing can now operate independently of any other, effectively independent asynchronous subroutines (whereas all early correlator designs were synchronous). The multiple “subroutines” can now implement end-to-end processing. These FPGAs are now available as COTS hardware ready to plug into a standard server and have 100GbE ports that can connect to the In-Network Processor.

The combination of these two technology advances allows a new approach to radio astronomy correlators and beamformers - what we have called “Atomic COTS”. This is a design that uses COTS In-Network Processors to route data to COTS FPGA cards such that each card receives part of the total bandwidth for all receiving elements. The FPGA then completes all processing to implement a correlator or beamformer independent of any other FPGA (an Atomic operation). The resulting astronomy data products can be routed back through the In-Network Processors to the next stage of processing such as imaging with visibilities and searches for FRBs on tied array beams. An Atomic COTS architecture will enable astronomy backends to upgrade and deploy incrementally.