Scaling the Fast x86 DVB-S2 Decoder to 1 Gbps

Software implementation of LDPC decoders has been an active area of development for the last 10 years. Researchers have focused on implementing the computationally expensive algorithm on both GPPs and GPUs. A major leap in performance was reported in the groundbreaking paper by Bertrand le Gal [2]. This paper builds on the work in [2] by considering the scaling of that implementation on modern many-core processors. We look at the performance of LDPC code specified in the DVB-S2 standard. The large block size of the DVB-S2 code makes the memory architecture of the processor just as important as the clock rate and instruction set. We present results for two generations of Intel Xeons, an Intel Phi (KNL), the recently released AMD EPYC. The key finding is that performance scaling is limited by the amount of available cache memory rather than the number of cores. We also find that heavily multi-threaded, but deterministic software architecture benefits from explicit allocation of threads to cores vs. allowing the operating system to manage threading. The maximum throughput of 1 Gbps was achieved on a mid-range AMD server - issuing a new era of all-software receivers for very high rate waveforms. We also present the performance of the algorithm ported to a low-power ARM processor and compare that to a low-end Intel Core.