Abstract
對於有限場倒數運算心縮陣列，我們提出了一種同步錯誤偵測法，可以在 系統執行有限場倒數運算的同時，偵測出系統本身是否有故障發生，用來 提高系統的可靠度。我們的方法不會影響系統速度上的表現，而且隨著系 統的位元數增加，為了同步錯誤偵測所多需多付出代價相對減少。我們的 方法屬於algorithm-based fault tolerance方法---在運算的路徑上，修 改原來的演算法，推論出在輸出和輸出之間parity的關係，使得同步錯誤 偵測可以借由檢測輸出和輸出之間parity的關係是否相符來決定。在控制 信號的路徑上，我們利用結構上的特性,觀察控制信號的兩端來決定是否 故障發生。同時觀察運算路徑上和控制路徑上的信號我們可以偵測陣列□ 所有的單一元素故障所發生的錯誤輸出。我們也用電腦模擬驗證我們的方 法，確認我們的方法可以偵測出所有的single stuck-at fault。當輸入 大於12個位元，我們所需多付出的晶片面積便小於5%。 A concurrent error detection method has been proposed for $GF(2 ^m)$ inversion systolic array. This method tests the circuit concurrently while it is in normal operation to increase the reliability of the system. There is neglectable performance penalty. Furthermore, the area overhead decreases as the input word size grows. This is achieved by an algorithm-based fault tolerance method, manipulating the Euclid's algorithm and defining the parity so that the resulting parity is the transformation of the parity at the inputs. Higher fault coverage can be achieved if we utilize the structure of the array to encode the control lines and other constant parameters. Using the above two techniques, analysis concludes that all single cell fault in the array can be detected concurrently. Simulation shows that 100% single stuck-at fault coverage can be achieved while the area overhead is smaller than 5% if the word- length is more than 12 bits.