Fire-Flyer 2

Fire-Flyer 2 is the second-generation AI-HPC cluster built by High-Flyer, a Chinese quantitative hedge fund, to train the deep-learning models behind its trading strategies; the same infrastructure and software stack later underpinned the DeepSeek language-model lab that High-Flyer spun out. High-Flyer's own research team published a detailed account accepted at SC24 ('Fire-Flyer AI-HPC: A Cost-Effective Software-Hardware Co-Design for Deep Learning', arXiv:2408.14158): 10,000 PCIe A100 40GB GPUs across roughly 1,250 eight-GPU nodes, a two-layer fat-tree InfiniBand fabric built from 122 Mellanox QM8700 200Gbps switches (versus the 1,320 switches an equivalent three-layer DGX-A100 topology would need) with one 200Gbps ConnectX-6 NIC per node, and a 3FS parallel filesystem across 180 storage nodes with over 20 PiB of NVMe capacity. The paper states total power draw of roughly 3+ MW for the GPU nodes, under 4 MW including network switches; because it gives a range rather than a single figure, power_kw is left blank here rather than picking a point estimate. High-Flyer's own company history page corroborates the 10,000-GPU figure and a stated 1 billion yuan cost, and dates the cluster to 2021, following the smaller, since-retired Fire-Flyer 1 (2020, roughly 1,100 GPUs; not recorded as a separate row here given inconsistent GPU counts for it across sources). No TOP500/HPL submission exists for this cluster.