AI Based Methods for Matrix Multiplication in High Resolution Simulations of Radio Access Networks

Abstract: The increasing demand for mobile data has placed significant strain on radio access networks (RANs), leading to a continuous need for increased network capacity. In keeping with that, a significant advancement in modern RANs is the ability to utilize several receivers and transmitters, to allow for beamforming. One way to increase the capacity of the network is therefore to optimize the resource allocation by preprocessing the transmitted signals, which involves several costly matrix multiplications (MMs). The aim of the project was to investigate the potential of accelerating Ericsson's RAN simulations by using AI based approximate matrix multiplication (AMM) algorithms. The main focus was on the multiply additionless (MADDNESS) algorithm, a product quantization technique that has achieved speedups of up to 100 times compared to exact MM, and 10 times faster than previous AMM methods. A complex matrix handling version of MADDNESS was implemented in Java and Python respectively, and its speed and accuracy were evaluated against Ericsson's current MM implementation. The proposed implementation did not beat the benchmark with respect to speed, instead resulting in a 4-10 times slowdown in runtime. However, this may largely be due to the fact that the used languages do not allow for complete control over memory resource allocation. As such, the implementations at hand do not incorporate all the crucial features of the algorithm. Particularly, the handicapped version does not fully leverage the vectorization potential, which is one of the key contributors to the speed of the algorithm. Consequently, further improvements are necessary before employing the techniques in an end-to-end implementation.