AI Paper Atlas

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

TPAMI 2024 · Large language models › Reasoning and chain of thought · approach: Efficient LLM serving · 16 citations (OpenAlex)

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 changedWhereMeasuredEffect
run BERT-Base on the memristor crossbar with 8-bit quantization and device noiseBERT-Base · GLUEaccuracy: SST-2 93.00 -> 93.12; MNLI 83.68 -> 83.77No clear effect
run BERT-Base on the memristor crossbarBERT-Base · GLUE RTEaccuracy: 65.34 -> 62.45Hurt
run BERT-Large on the memristor crossbarBERT-Large · GLUE CoLAMatthews corr.: 62.63 -> 62.43No clear effect
deploy Phi-1.5 on the crossbar with 4-cycle computationPhi-1.5 · MMLUaccuracy: 0.418 -> 0.398Hurt
deploy Phi-1.5 on the crossbar with 3-cycle computation instead of 4-cyclePhi-1.5 · WinoGrandeaccuracy: 0.729 -> 0.717Hurt
use 3-cycle instead of 4-cycle computationGPT-2 · ARC_Easyaccuracy: 0.439 -> 0.406Hurt
use 3-cycle instead of 4-cycle computationGPT-2 · PIQAaccuracy: 0.637 -> 0.599Hurt
choose encoding base 7 (S=2) rather than base 15 or larger (ablation)computation crossbar · INT8 activation encodingscale-cycle product: 6 vs 9, 8, 10, 12Improved
choose scale factor S=2 rather than S=1 (ablation)computation crossbar · INT8 activation encodingarea-delay product: about 9% reduction in ADPImproved
replace multi-bit crossbar with dense plus computation crossbar designmemristor architecture · area overheadarea saving: 6x savings vs multi-bit; 39x vs single-bitImproved
replace conventional crossbars with dense plus computation crossbar designmemristor architecture · energy consumptionenergy reduction: 18x vs multi-bit; 3x vs single-bitImproved
amplify device noise 5x (ablation)dense crossbar · model accuracyaccuracy: noise amplitude 5X strongerNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy