https://hgpu.org/?p=16050
Persistent RNNs: Stashing Recurrent Weights On-Chip