Matrix compression methods

Abstract

The biggest cost of computing with large matrices in any modern computer is related to memory latency and bandwidth. The average latency of modern RAM reads is 150 times greater than a clock step of the processor (Alted, 2010). Throughput is a little better but still 25 times slower than the CPU can consume. The application of bitstring compression allows for larger matrices to be moved entirely to the cache memory of the computer, which has much better latency and bandwidth (average latency of L1 cache is 3 to 4 clock steps). This allows for massive performance gains as well as the ability to simulate much larger models efficiently. In this work, we propose a methodology to compress matrices in such a way that they retain their mathematical properties. Considerable compression of the data is also achieved in the process. Thus allowing for the computation of much larger linear problems within the same memory constraints when compared with the traditional representation of matrices.