Associative Memristive Memory for Approximate Computing in GPUs

Associative memory, in the form of lookup tables, is a promising approach to improving energy efficiency by enabling computing-with-memory. A processing element can be tightly coupled with an associative memory where function responses are pre-stored. Associative memories can recall function responses for a subset of input values therefore avoiding the actual function execution on the processing element that leads to energy saving. One challenge however is to reduce the energy consumption of associative memory modules themselves. In this paper, we address the challenge of designing ultra-lowpower associative memories. We first use memristive parts for memory implementation and demonstrate the energy saving potential of integrating associative memristive memory (AMM) into graphics processing units (GPUs). Next, we leverage approximate computing which takes advantage of application-level tolerance to errors, to enable voltage overscaling to further reduce energy consumption of an AMM module. Voltage overscaling deliberately relaxes the searching criteria of an AMM: The AMM module finds stored patterns matching an input search pattern with a Hamming distance of 0, 1, or 2. This controllable inexact matching introduces some errors to the computation, that are tolerable for the target application. The energy consumption is further reduced by employing a purely resistive crossbar architecture for the AMM module. To evaluate our solution, we tightly integrate AMM modules with floating point units (FPUs) in an AMD Southern Islands GPU. Then we run four image processing kernels on an AMM-integrated GPU to evaluate the proposed architecture. Our experimental results show that the use of the AMM modules reduces energy consumption of running these kernels on GPU by, on average, 23%–45%, compared to the baseline GPU without AMM modules. We also show that these image processing kernels can tolerate errors resulting from approximate search operations with an acceptable degradation of image quality, i.e., a PSNR greater than 30dB.