With the exponential expansion of large-scale datasets, there has been a burgeoning exploration of distributed computing techniques within the realm of machine learning. Among these, the extension of large-scale stochastic gradient descent (SGD) algorithms has emerged as particularly promising due to its notable efficiency in handling vast datasets. In response to the challenges inherent in distributed computing and the imperative to enhance model efficiency, this paper introduces an innovative algorithm termed local update SGD. This algorithm is further augmented by integrating it with the quantized SGD (QSGD) technique, resulting in the proposal of a novel hybrid algorithm. Through a meticulous comparison and analysis of four distinct algorithms using the MNIST dataset, this new hybrid algorithm demonstrates a substantial reduction in training time compared to synchronous SGD, while upholding model accuracy. Notably, the hybrid algorithm exhibits only marginal improvements in training time when juxtaposed with local update SGD.