AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
AI Chips & Compute

Zhongcheng Hualong Unveils HL200 Inference Chip and 10,000-Card Supernode Cluster

Zhongcheng Hualong launches HL200 inference chip and supernode cluster for up to 10,240 GPUs.

AI News Feed · 2026-08-25 15:37

The launch came as the AI industry shifts from training to inference. According to the company, the rollout of AI agents, long-context interactions and multi-turn dialogues has pushed up token consumption, making low-latency inference for ultra-long contexts a core competitive advantage.

Built on previous-generation technology, HL200 natively supports FP4 and FP8 ultra-low-precision inference as well as mainstream FP16. The company said single-card computing power reaches 0.5 P for FP16/BF16, 2 P for FP8 and 4 P for FP4, with energy efficiency of 5.12 TFLOPS/W. The chip is compatible with PyTorch, ONNX and vLLM, and offers an OpenAI-compatible interface.

In benchmark tests against international chips of the same class, HL200 showed leading performance in the full Prefill pipeline and lower decoding latency on models including DeepSeek V4 Flash, DeepSeek V4 Pro and GLM 5.2, the company said.

The super-node cluster supports 64 GPUs per cabinet, up to 1,024 GPUs stacked vertically per node, and up to 10,240 GPUs horizontally, covering deployments from 8 cards to more than 10,000 cards. The company said the system uses computing, storage and networking SLO coordination to improve density, bandwidth, latency and energy efficiency.

Bi Kaichun, executive deputy director and secretary-general of the Electronic Information Science and Technology Committee of the Ministry of Industry and Information Technology, said at the event that domestic AI chips have become a strategic necessity. He called on Zhongcheng Hualong to continue research into core technologies and improve ecosystem adaptation.

Liu Zhiguang, chairman of the China Science and Technology Consulting Association, said the national 15th Five-Year Plan encourages enterprises to lead key technology research and innovation alliances. He called the HL series an example of self-reliant computing power development.

The company also signed strategic cooperation agreements with more than ten enterprises, including China Energy Engineering's Zhongdian Engineering, Inspur Information, Unisplendour Smart Computing and Sinnet Cloud. Wang Jiacheng, chairman of Zhongcheng Hualong, said computing power competition has entered a stage of comprehensive competition in efficiency, deployment experience and industrial value.

Zhongcheng Hualong described itself as one of the few domestic companies with full-stack capability covering chips, systems and solutions. It said its products have won bids from China Mobile and China Telecom and have been selected in nearly 50 key sections in more than 20 provinces.