Librería Samer Atenea
Kálamo Books
Librería Elías (Asturias)
Librería Kolima (Madrid)
Librería Proteo (Málaga)
Thisbook tackles a critical bottleneck in large-scale AI: the slow and communication-heavy training of massive Deep Neural Networks (DNNs) on multi-GPU systems. It addresses the trade-off between two main parallelization methods. Data parallelism suffers from severe communication overhead for large models, while pipelined model parallelism (like PipeDream) offers up to 8.91x speedup for large Fully Connected/Recurrent Neural Networks but causes 'weight staleness,' degrading model accuracy. To resolve this, the paper introduces SpecTrain, a novel technique. SpecTrain uses the momentum from optimizers to predict future weight updates, allowing pipelined computation with accurate, non-stale weights. This enables the high GPU utilization and speed of pipelining while maintaining the training robustness and final accuracy of synchronous methods.
