Research map / Architectures and efficiency / RNNs and normalization
Sparse Persistent RNNs: Squeezing Large Recurrent Networks On-Chip
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use sparse persistent RNNs instead of dense GEMM | RNN layer size 1792, batch size 4, density 10% · RNN, layer size 1792, batch size 4, density 10% | speedup: 7.3× | Helped |
| use sparse persistent RNNs instead of sparse GEMM | RNN layer size 1792, batch size 4, density 10% · RNN, layer size 1792, batch size 4, density 10% | speedup: 3.4× | Helped |
| use sparse persistent RNNs instead of dense persistent RNNs | RNN layer size 1792, batch size 4, density 10% · RNN, layer size 1792, batch size 4, density 10% | speedup: 1.8× | Helped |
| use sparse persistent RNNs instead of dense GEMM | RNN layer size 2304, density 10% · RNN, layer size 2304, density 10% | speedup: 9.9× | Helped |
| use sparse persistent RNNs instead of dense GEMM | RNN layer size 5760, density 1% · RNN, layer size 5760, density 1% | speedup: 10.6× | Helped |
| use sparse persistent RNNs instead of sparse GEMM | RNN layer size 5760, density 5% · RNN, layer size 5760, density 5% | performance | Hurt |
| add wide memory loads (ablation) | Sparse persistent RNN · Layer size 1152, batch size 4, density 10%, 256 timesteps | speedup vs. dense GEMM: 2.53× -> 4.28× | Helped |
| add bank-aware weight layout (ablation) | Sparse persistent RNN with wide memory loads · Layer size 1152, batch size 4, density 10%, 256 timesteps | speedup vs. dense GEMM: 4.28× -> 4.56× | Helped |
| add Lamport timestamps (ablation) | Sparse persistent RNN with bank-aware layout · Layer size 1152, batch size 4, density 10%, 256 timesteps | speedup vs. dense GEMM: 4.56× -> 5.44× | Helped |
| increase density from 10% to 1% | Sparse persistent RNN · RNN, layer size 1792 | speedup vs. dense GEMM: 7.3× -> 15.0× | Helped |
| use row-balanced rather than naïve pruning (ablation) | NMT LSTM · NMT validation | BLEU: less than 0.1 BLUE points | No clear effect |
| use load-balanced pruning instead of straightforward unbalanced pruning (ablation) | RNN layer size 2304, batch size 4, density 25% · RNN, layer size 2304, density 25% | speedup over dense GEMM: 1.14× -> 1.94× | Helped |
Closest papers by meaning
- Persistent RNNs: Stashing Recurrent Weights On-Chip (ICML 2016 · similarity 0.94)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.91)
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.91)
- Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.89)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.87)
- Harnessing Manycore Processors with Distributed Memory for Accelerated Training of Sparse and Recurrent Models (AAAI 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.