单芯片到万卡集群体系化突破 中诚华隆HL200推理芯片及超节点集群重磅发布

TL;DR - Zhongcheng Hualong launched its second-generation HL200 AI inference chip and a supernode cluster architecture scaling to 10,240 GPUs. The offering targets high-throughput, low-latency deployment of trillion-parameter models and agent workloads on domestically produced Chinese infrastructure.

• HL200 supports FP4, FP8, FP16, and BF16 inference, with claimed per-card performance of 4 PFLOPS at FP4, 2 PFLOPS at FP8, and 0.5 PFLOPS at FP16/BF16.
• The company reports energy efficiency of 5.12 TFLOPS/W and says internal comparisons show strong prefill performance and low decoding latency, though no independent benchmarks are provided.
• Its software stack supports PyTorch, ONNX, vLLM, and OpenAI-compatible APIs to reduce model-migration costs.
• The cluster design connects 64 GPUs per rack, scales vertically to 1,024 GPUs per node, and expands horizontally to 10,240 GPUs.