Zhongcheng Hualong Releases HL200 AI Inference Chip with 5.12 TFLOPS/W Efficiency
Chinese hardware firm Zhongcheng Hualong has released the HL200 AI inference chip, claiming an energy efficiency of 5.12 TFLOPS/W alongside native support for low-precision FP4 and FP8 formats. The company also introduced a cluster solution capable of scaling up to 10,240 cards. This release represents a significant step in China's domestic AI hardware ecosystem, aiming to reduce reliance on foreign chips by offering high-efficiency inference capabilities. The support for massive clustering addresses the growing demand for scaling large language models and complex AI workloads. The HL200 delivers 4 PFLOPS of FP4, 2 PFLOPS of FP8, and 0.5 PFLOPS of FP16/BF16 computing power per card. Its super-node cluster solution supports 64 GPUs per cabinet and allows vertical stacking of up to 1,024 cards.
## BACKGROUND
AI inference chips are specialized processors optimized for running trained machine learning models rather than training them. Low-precision formats like FP4 (4-bit floating point) are increasingly adopted in modern AI hardware to accelerate inference speeds, reduce memory usage, and improve energy efficiency.