Research map / Architectures and efficiency / Mixture of experts
GShard: Scaling Giant Models with Conditional Computation and Automatic Sharding
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 |
|---|---|---|---|
| Scale the MoE Transformer to 2048 experts per layer and 36 layers | MoE Transformer · M4, 100 languages to English | average BLEU: 44.3 vs 36.9 | Helped |
| Increase depth from 12 to 36 layers while keeping experts per layer fixed (ablation) | MoE Transformer · M4, 100 languages to English | average BLEU: 2-to-3 BLEU points on average | Helped |
| Increase experts per layer from 128 to 512 at 12 layers (ablation) | MoE Transformer · M4, 100 languages to English | average BLEU: +3.3 average BLEU | Helped |
| Increase experts per layer from 512 to 2048 at 12 layers (ablation) | MoE Transformer · M4, 100 languages to English | average BLEU: +1.3 average BLEU | Helped |
| Increase expert count while fixing depth (ablation) | MoE Transformer · High-resource languages in M4 | BLEU | Helped |
| Increase expert count while fixing depth (ablation) | MoE Transformer · Low-resource languages in M4 | BLEU | Mixed |
| Increase depth from 12 to 36 layers (ablation) | MoE Transformer · M4 training convergence | billion tokens to reach cross-entropy 0.7: 995 -> 321 | Helped |
| Use a 36-layer MoE Transformer rather than a 12-layer model with the same expert count (ablation) | MoE Transformer · M4 training convergence | tokens to preset training loss: 2 to 3 times fewer tokens | Helped |
| Use conditional MoE computation instead of dense scaling | Multilingual Transformer · M4 training | TPU core years: 22.4 vs $\sim$235.5 | Helped |
| Increase expert count from 128 to 2048 (ablation) | MoE Transformer · TPU runtime scalability | device time: 1.7x execution time increase when ... scale the model by 16x | Hurt |
Closest papers by meaning
- Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 2025 · similarity 0.86)
- The Journey Matters: Average Parameter Count over Pre-training Unifies Sparse and Dense Scaling Laws (ICLR 2025 · similarity 0.86)
- Pruning Large Language Models with Semi-Structural Adaptive Sparse Training (AAAI 2025 · similarity 0.86)
- Search for Efficient Large Language Models (NeurIPS 2024 · similarity 0.86)
- GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling (NeurIPS 2025 · similarity 0.86)
- Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.86)
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.