Ideally, to validate the effectiveness of COAL, we have to simulate the entire design space consisting of 70 million design configurations, then compare the configuration deduced by COAL and the actual optimal configuration. However, this is infeasible due to intractably large simulation costs. A compromise is made by comparing the promising configuration deduced by COAL with the configuration deduced by the state-of-the-art DSE approach (ANN). To be specific, we simulate the two architectural configurations on two illustrative benchmarks, and compare the corresponding processor responses (performance and power) directly.
As an example, we employ the benchmarks bzip2 and crafty in the experiments. Suppose the design specification is that “maximizing the performance with the constraint that the power consumption must be less than 100 watt”, we can attain the promising
Ideally, to validate the effectiveness of COAL, we have to simulate the entire design space consisting of 70 million design configurations, then compare the configuration deduced by COAL and the actual optimal configuration. However, this is infeasible due to intractably large simulation costs. A compromise is made by comparing the promising configuration deduced by COAL with the configuration deduced by the state-of-the-art DSE approach (ANN). To be specific, we simulate the two architectural configurations on two illustrative benchmarks, and compare the corresponding processor responses (performance and power) directly.
As an example, we employ the benchmarks bzip2 and crafty in the experiments. Suppose the design specification is that “maximizing the performance with the constraint that the power consumption must be less than 100 watt”, we can attain the promising
การแปล กรุณารอสักครู่..
