Design space exploration in many-core processors for sound synthesis of plucked string instruments

Recent advances in physics-based sound synthesis have unveiled numerous possibilities for the creation of new musical instruments. Despite the fact that research on physics-based sound synthesis has been going on for three decades, its higher computational complexity compared to that of signal modeling has limited its use in real-time applications. This limitation has motivated research on parallel processing architectures that support the physics-based sound synthesis of musical instruments. In this paper, we present analytical results of the design space exploration of many-core processors for the physics-based sound synthesis of plucked-string instruments including acoustic guitar, classical guitar and the gayageum, which is representative of a Korean plucked-string instrument. We do so by quantitatively evaluating the significance of a sample-per-processing-element (SPE) ratio–i.e., the amount of sample data directly mapped to each processing element, which is equivalent to varying the number of processing elements for a fixed sample size on system performance and efficiency using architectural and workload simulations. The effect of the sample-to-processor ratio is difficult to analyze because it fundamentally affects both hardware and software design when varied. In addition, the optimal SPE ratio is not typically at either extreme of its range–i.e., one sample per processor or one processor per an entire sample. This paper illustrates the correlation between a fixed problem sample size, SPE ratio and processing element (PE) architecture for a target implementation in 130-nm CMOS technology. Experimental results indicate that an SPE in the range of 5513 to 2756, which is equivalent to 48 to 96 PEs for guitars and 96 to 192 PEs for the gayageum, provides the most efficient operation for the synthesis of musical sounds sampled at 44.1 kHz, yielding the highest task throughput per unit area or per unit energy. In addition, the produced synthesized sounds appear to be very similar to the original sounds, and the selected optimal many-core configurations outperform commercial processor architectures including DSPs, FPGAs, and GPUs in terms of area efficiency and energy efficiency.

► We explore the design space of many-core processors for sound synthesis algorithms.
► We quantitatively evaluate the significance of a sample-per-processing-element (SPE) ratio.
► We illustrate the correlation between a fixed problem sample size, SPE ratio and PE architecture.
► An SPE in the range of 5513 to 2756 provides the most efficient operation.