Bank-interleaved cache or memory indexing does not require euclidean division
Abstract
! Abstract Concurrent access to bank-interleaved memory structure have been studied for decades, particularly in the context of vector supercomputer systems. It is still common belief that using a number of banks different from 2 n leads to insert a complex hardware including a non-trivial divider on the access path to the memory. In 1993, two independent studies [1], [2] were showing that through leveraging a very simple arithmetic result, the Chinese Remainder Theorem, this euclidean division is not needed when the number of banks is prime or simply odd. In the mid 90's, the interest for vector supercomputers faded and the research topic disappeared. The interest for bank-interleaved cache has reappeared recently [3] in the GPU context. In this short paper, we extend the result from [1] and we show that, regardless the number of banks: Bank-interleaved cache or memory indexing does not require euclidean division.
Domains
Hardware Architecture [cs.AR]
Origin : Files produced by the author(s)
Loading...