Hanguang 800 NPU (AI Inference Chip)

Unveiled at the Apsara Conference on 2019-09-25 by T-Head under Alibaba DAMO Academy. It uses a self-developed hardware framework with algorithms optimized for retail and logistics workloads in the Alibaba ecosystem. At launch it was already used internally for e-commerce product search, automatic translation, personalized recommendation, advertising and intelligent customer service. Alibaba stated it plans to empower clients by providing access through its cloud business (no standalone hardware sales). Official ResNet-50 inference results: single-chip peak 78,563 IPS; energy efficiency 500...

Key Highlights

• Self-developed hardware architecture optimized for cloud AI inference
• ResNet-50 peak 78,563 IPS and 500 IPS/W energy efficiency (official press release)
• powers internal e-commerce search/recommendation/ads/customer service
• compute delivered through Alibaba Cloud services

Technical Specifications

Typical Use Cases

仅通过云服务提供，不单独发售硬件（官方新闻稿原文：plan to empower clients by providing access through our cloud business）；官网无独立在售加速卡产品页，无公开硬件 datasheet，官方新闻稿页面未提供产品图片资源，故 image_local_paths / image_source_urls 留空；category 四值中无“云端 NPU 芯片”对应项，取最接近的 devkit_or_module（芯片级产品），实际交付形态为阿里云云端算力实例；官方公开数据仅 ResNet-50 两项指标，其余硬件规格缺失，按规范留空不补；研发主体为达摩院平头哥 T-Head。