Research map / Large language models / Reasoning and chain of thought
Enabling Energy-Efficient Deployment of Large Language Models on Memristor Crossbar: A Synergy of Large and Small
Zhehui Wang, Tao Luo, Cheng Liu, Weichen Liu, Rick Siow Mong Goh, Weng‐Fai Wong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| run BERT-Base on the memristor crossbar with 8-bit quantization and device noise | BERT-Base · GLUE | accuracy: SST-2 93.00 -> 93.12; MNLI 83.68 -> 83.77 | No clear effect |
| run BERT-Base on the memristor crossbar | BERT-Base · GLUE RTE | accuracy: 65.34 -> 62.45 | Hurt |
| run BERT-Large on the memristor crossbar | BERT-Large · GLUE CoLA | Matthews corr.: 62.63 -> 62.43 | No clear effect |
| deploy Phi-1.5 on the crossbar with 4-cycle computation | Phi-1.5 · MMLU | accuracy: 0.418 -> 0.398 | Hurt |
| deploy Phi-1.5 on the crossbar with 3-cycle computation instead of 4-cycle | Phi-1.5 · WinoGrande | accuracy: 0.729 -> 0.717 | Hurt |
| use 3-cycle instead of 4-cycle computation | GPT-2 · ARC_Easy | accuracy: 0.439 -> 0.406 | Hurt |
| use 3-cycle instead of 4-cycle computation | GPT-2 · PIQA | accuracy: 0.637 -> 0.599 | Hurt |
| choose encoding base 7 (S=2) rather than base 15 or larger (ablation) | computation crossbar · INT8 activation encoding | scale-cycle product: 6 vs 9, 8, 10, 12 | Improved |
| choose scale factor S=2 rather than S=1 (ablation) | computation crossbar · INT8 activation encoding | area-delay product: about 9% reduction in ADP | Improved |
| replace multi-bit crossbar with dense plus computation crossbar design | memristor architecture · area overhead | area saving: 6x savings vs multi-bit; 39x vs single-bit | Improved |
| replace conventional crossbars with dense plus computation crossbar design | memristor architecture · energy consumption | energy reduction: 18x vs multi-bit; 3x vs single-bit | Improved |
| amplify device noise 5x (ablation) | dense crossbar · model accuracy | accuracy: noise amplitude 5X stronger | No clear effect |
Closest papers by meaning
- xLSTM: Extended Long Short-Term Memory (NeurIPS 2024 · similarity 0.84)
- Analog Foundation Models (NeurIPS 2025 · similarity 0.83)
- NoMAD-Attention: Efficient LLM Inference on CPUs Through Multiply-add-free Attention (NeurIPS 2024 · similarity 0.83)
- SpeedLoader: An I/O efficient scheme for heterogeneous and distributed LLM operation (NeurIPS 2024 · similarity 0.83)
- ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization (NeurIPS 2024 · similarity 0.83)
- Nemotron-Flash: Towards Latency-Optimal Hybrid Small Language Models (NeurIPS 2025 · similarity 0.83)
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.
