Research map / Architectures and efficiency / RNNs and normalization
Persistent RNNs: Stashing Recurrent Weights 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 |
|---|---|---|---|
| keep recurrent weights in on-chip registers and reuse across timesteps (persistent kernels) instead of GEMM | deep RNN at mini-batch 4 · TitanX, layer size 1152 | TFLOP/s: 0.099 to 2.8 (~30x) | Helped |
| add residual skip connections between RNN layers (ablation) | 48-layer RNN speech model · 500h speech dev set | Dev WER: 100.0 without skip to 29.40 with skip 4 | Helped |
| skip every 4 layers vs every 1 layer (ablation) | 48-layer RNN speech model · 500h speech dev set | Dev WER: 38.77 (skip 1) vs 29.40 (skip 4); skip 8 gives 27.44 but authors cannot explain it | Mixed |
| increase depth at constant total parameters | RNN speech model · 500h speech dev set | Dev WER: 33.61 at 8 RNN to 31.64 at 40 RNN, 28.12 at 56, then 30.67 at 72 and 42.67 at 88 | Mixed |
| increase depth by stacking layers (constant params per layer) | RNN speech model · 500h speech dev set | Dev WER: 35.69 (8 RNN) to 27.23 (88 RNN, 234M params) | Helped |
| increase mini-batch size beyond 512-1024 for training | RNN speech model · 500h speech dev set | epochs to convergence | Hurt |
Closest papers by meaning
- FlashRNN: I/O-Aware Optimization of Traditional RNNs on modern hardware (ICLR 2025 · similarity 0.88)
- Accelerating Linear Recurrent Neural Networks for the Edge with Unstructured Sparsity (ICML 2025 · similarity 0.88)
- Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training (AAAI 2024 · similarity 0.86)
- Harnessing Manycore Processors with Distributed Memory for Accelerated Training of Sparse and Recurrent Models (AAAI 2024 · similarity 0.86)
- Recurrent neural networks: vanishing and exploding gradients are not the end of the story (NeurIPS 2024 · similarity 0.85)
- M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling (NeurIPS 2026 · similarity 0.85)
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.