A Case Study on the Design Trade-off of a Thread Level Data Flow based Many-core Architecture

With the potential of overcoming the memory and power wall, the many-core/multi-thread has become a trend in processor design area. However, this architecture is far from ripeness because it also companies with many challenges such as scalability and larger architecture design space compared with mono-core architectures. In many-core design space, Data-Flow based architectures are alternatives that deal with concurrency, long memory latencies, and synchronization stalls efficiently. Nevertheless, even in this sub-area, there are still a lot of factors affecting the scalability and performance of the architecture. In this paper, we explore the design trade-offs for Decoupled Threaded Architecture (DTA) which is a data-flow many-core architecture. By using a well known bio-informatics benchmark, ClustalW, we evaluate various DTA configurations with different number of synchronization and execution pipelines. We find that the configuration which consists of two synchronization pipelines (SP) and one execution pipeline (EP) for each processing element(PE) achieves almost the same performance as the configuration consisting of two SPs and two EPs for each processing element. By employing the former configuration, we can save 32.5%% of the area required for each DTA processing element