Insights into DeepSeek-V3: hardware co-design paper
DeepSeek lays out the hardware limits it hit training V3 on 2,048 H800s and argues for low-precision units and better interconnects.
Reflection paper on MLA, MoE, FP8 training and a multi-plane network topology, framing V3's design as hardware-aware co-design under export-controlled GPUs, with proposals for future accelerators.
- Date
- Wednesday, 14 May 2025
- Lab
- DeepSeek
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Training cluster | 2,048 NVIDIA H800 GPUs V3 training | company |
Sources
This record was checked against its sources on 6 October 2026. How we check