面向多处理器SoC设计的低功耗软硬件划分

Low power hardware/software partitioning for multiprocessor SoC design

提出了解决多处理器SoC的低功耗软硬件划分问题的方法——基于神经网络的禁忌搜索算法。其基本思想是：真实的生物神经元具有抑制重复激活的阻尼特性,这与禁忌搜索对重复搜索加以限制相类似,因此设计具有阻尼特性的神经网络实现禁忌搜索算法,受阻尼特性抑制的神经元对应禁忌活动。由于神经网络复杂的动态特性和禁忌搜索优秀的全局搜索能力,该算法能够有效地跳出局部最优解。对真实任务图的实验表明,与遗传算法相比,该算法不但具有搜索速度上的优势,而且所得到的绝大部分软硬件划分方案有更低的系统功耗。

An algorithm based on tabu search on a neural network was put forward to solve the low power hardware/software partitioning problem in design of system on chip （SoC） architectures consisting of several types of processors. The basic idea of it is： the refractory effect of inhibiting the repetitive firings is one of the characteristics of real biological neurons, which is similar to the tabu effect of tabu search, so tabu search can be realized by a neural network in which neurons inhibited by the refractory effect correspond to the tabu moves. With the complex dynamics of neural networks and excellent global search capacity of tabu search, the algorithm can effectively avoid trapping in undesirable local minima. The experiments for real task graphs show that the algorithm has better time performance than the genetic algorithm, and most of hardware/software partitioning solutions gotten from the algorithm possess the lower power consumption.