Understanding Throughput-Oriented Architectures

The fundamental throughput-oriented architectural design decisions differentiating throughput-oriented processors from more traditional latency-oriented architectures are explored by using GPUs. Throughput-oriented processors rely on three key architectural features such as emphasis on many simple processing cores, extensive hardware multi-threading, and use of single-instruction, multiple-data, or SIMD, execution. These processors achieve even higher levels of performance by using many simple, and hence small, processing cores. The GPU employs a single-instruction, multiple-thread, or SIMT, architecture in which threads resident on a single SM are executed in groups of 32, called warps. GPUs are also a powerful platform for exploring scalable algorithms and a leading indicator for algorithm design on future throughput-oriented architectures.