Zhenwu M890 All-Scenario Training & Inference AI Chip

Product in T-Head's official AI-chip section. End-to-end self-developed hardware; upgraded self-developed parallel-compute architecture and on-chip interconnect. The self-developed T-Head SAIL® software stack has independent IP, is compatible with mainstream AI ecosystems, provides a unified programming API and requires no application-code modification for migration; supports multi-process resource sharing and multi-tenant isolation; chip compute is bundled with Alibaba AI frameworks, platforms, models and applications. Official specs: 144 GB memory; 800 GB/s inter-chip interconnect; FP8/FP4...

Key Highlights

• Fully self-developed
• high performance (144 GB memory, 800 GB/s interconnect, FP8/FP4, 64-card full-bandwidth mesh)
• easy to use (SAIL® full stack, unified API, zero-code migration)
• flexible deployment (multi-process sharing, tenant isolation)
• chip-cloud-model integration

Technical Specifications

Typical Use Cases

官网公开为芯片级规格（显存/互联带宽/精度），未公布板卡形态、TDP、制程工艺与公开 PDF datasheet（页面“完整 Datasheet”需商务合作/技术工单获取）；产品通过阿里云生态与合作渠道交付，非公开零售加速卡；图片为官网产品页主视觉渲染图（芯片 die 与数据中心机柜）；category 取 accelerator_card（AI 加速芯片产品线），实际交付形态以官方云实例/商务方案为准。本地图片为官网产品页主视觉的产品区域裁剪版（裁剪掉左侧空白），原始图见 image_source_urls。