Fire-Flyer AI-HPC: cost-effective cluster co-design
DeepSeek documents its 10,000-GPU PCIe A100 cluster, which it says nears DGX-A100 performance at half the cost and 40% less energy.
Describes the Fire-Flyer 2 design, which uses PCIe A100s rather than DGX nodes, HFReduce for allreduce, a congestion-free storage-compute network, and the HaiScale, 3FS and HAI-Platform software stack that later became the open 3FS.
- Date
- Monday, 26 August 2024
- Lab
- DeepSeek
- Kind
- paper
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Cluster cost vs DGX-A100 | about half 10,000 PCIe A100 GPUs; 40% less energy; performance approximating DGX-A100 (company-claimed) | company |
arXiv 2408.14158 submitted 2024-08-26. A first public account of the hardware behind DeepSeek's early models; the later V3 used H800s.
Sources
This record was checked against its sources on 6 October 2026. How we check