A 4D Hybrid Algorithm to Scale Parallel Training to Thousands of GPUs

Large communication costs are a critical bottleneck in training state-of-the-art neural networks on distributed systems. This paper introduces AxoNN, a novel four-dimensional (4D) parallelization approach, inspired by Agarwal's algorithm for matrix multiplication, for parallelizing tensor computations in deep learning, AxoNN employs two key strategies to minimize communication overhead. First, we optimize communication by overlapping expensive collective operations (reduce-scatter, all-gather, all-reduce) with computations. Our experiments with a 20-billion parameter transformer model demonstrate that these optimizations deliver nearly 53\% improvement. Second, we present an analytical model to assist users in identifying communication-minimizing configurations within the vast search space defined by our 4D algorithm. This model empowers practitioners by simplifying the tuning process for their specific training workloads. When training an 80-billion parameter model on 1024 GPUs of Perlmutter, AxoNN surpasses Megatron-LM, a state-of-the-art framework, by a significant 26%. Additionally, it achieves 57% of the theoretical peak FLOP/s.