Читать книгу Multi-Processor System-on-Chip 2 - Liliana Andrade - Страница 32

This chapter closely followed an SW implementation of the GFDM algorithm on the SotA vDSP and noted considerations taken into account with regard to handset workloads expected in modern and future mobile communications. We give analyses and conclusions on four layers: specification requirements, translating theory to pseudo-code, precision analysis and requirements and implementation space exploration.

First, we indicated trends and analyzed specifications for the most advanced 5G communication standard at the time of writing this chapter and defined workloads based on these specifications. To get closer to HW requirements, we selected an example beyond 5G algorithm GFDM to implement and investigate its HW requirements for those workloads. Second, from the theory behind GFDM, we developed the processing graph and pseudo-code that does not use modulo operations. Third, the precision analysis indicated that HW should be able to handle data types of at least 32 bits (16 bit – real, 16 bit – imag) per data point for operation and that the processor accumulators should have at least [log₂M] guard bits to facilitate computations without degradation. Fourth, we identified the range of kernel parameters as well as the number of calls and deadlines for the low- and high-end 5G workloads, followed by the implementation space exploration, resulting in high-throughput and low-memory access code variants of the same kernel. Finally, the kernel variants and workloads are combined to indicate required HW for corner cases. For the low-end and CA high-end vDSPs are recommended, while for higher than those and towards MIMO CA high-end workloads, more specialized (application-specific) HW accelerators or ASIPs are recommended.

We can see that giving a recommendation or direct HW requirements for SotA specifications is a multi-layered challenge that is difficult to overcome. For the overall system, other kernels can be analyzed in the same manner; however, the overarching theme is the need for variability and flexibility at each step of the analysis. Heterogeneous HW MPSoCs consisting of RISCs, vDSPs, ASIPs and dedicated (application-specific) HW accelerators can provide both flexibility and highly optimized solutions in the power consumption/chip area/chip cost sense, provided that enough care and effort has been invested into the investigation of the above-mentioned layers.