Block-sparse GPU kernels

Researchers released optimized GPU kernels for block-sparse neural networks, speeding up sparse linear ops in fully connected and convolutional layers.

The kernels support TensorFlow and can work with models like LSTMs, aiming to make sparsity practical on GPUs.

They report state-of-the-art results on some sentiment analysis and generative modeling tasks, including runs on an NVIDIA Titan X Pascal.

By sharply reducing compute and memory needs, the approach could enable much larger models and new sparsity-based architectures.