Fast parallel prefix logic circuits for n2n round-robin arbitration

Abstract

An n2n round-robin arbiter (RRA) searches its n inputs for a 1, starting from the highest-priority input. It picks the first 1 and outputs its index in one-hot encoding. RRA aims to be fair to its inputs and maintains fairness by simply rotating the input priorities, i.e., the last arbitrated input becomes the lowest-priority input. Arbiters are used to multiplex the usage of shared resources among requestors as well as in dispatch logic where the purpose is load balancing among multiple resources. Today, arbiters have hundreds of ports and usually need to run at very high clock speeds. This article presents a new gate-level RRA circuit called Thermo Coded-Parallel Prefix Arbiter (TC-PPA) that scales to any number of requestors. It uses parallel prefix network topologies (borrowed from fast carry lookahead adders) to generate a thermometer-coded pointer, thus greatly reducing critical path. Code generators were written not only for TC-PPA but also for the 5 highly competitive circuits in the literature (9 including their variants), and a rich set of timing/area results were obtained using a standard-cell based logic synthesis flow with a novel iterative strategy based on binary search. Synthesis runs include results with wire-load and without. Results show that for 54 or more ports (except 256) TC-PPA offers the best timing (lowest latency) as well as competitive area. Contributions also include transaction-level simulations that show when pipelining is used to boost clock rate, latency and input FIFO sizes are adversely affected, and hence pipelining cannot be indiscriminately exploited to trim clock period. Graphical abstract [Display omitted] Highlights► Thermo Coded-Parallel Prefix Arbiter (TC-PPA) circuit topologies are proposed. ► TC-PPA offers the fastest Round-Robin Arbiter circuit for 54 ports or more. ► TC-PPA owes its speed in part to Parallel Prefix networks found in fast adders. ► Verilog generators were written and synthesis was run for 13 circuit topologies. ► Results with up to 512 ports, wire-load model, and pipelining are included.