Atlas 850E

The Atlas 850E air-cooled supernode uses the Ascend 950DT, with 8 NPUs per compute node delivering up to 14.2 PFLOPS @mxFP4 and up to 4.0 TB/s read/write bandwidth. Through the Lingqu 2.0 protocol it provides 13.4TB/s aggregate NPU interconnect bandwidth per frame, fits standard air-cooled data centers, and serves internet, finance, government and education AI workloads.

Key Highlights

• Low-precision inference: industry-leading 14.2 PFLOPS @mxFP4 with HiF8/mxFP8/mxFP4. Ultra-high memory speed: industry-leading 4.0 TB/s HBM bandwidth, >=768GB HBM per system. Ultra-high interconnect: up to 13.4 TB/s aggregate NPU interconnect
• composable into 96- or 768-card air-cooled AI supernodes over Lingqu.

Technical Specifications

Typical Use Cases

归入supernode_cluster（风冷超节点，官网路径/hardware/cluster?tag=850e）；英文站无对应英文产品页，英文文案为翻译稿；img1: 视角1官方渲染图（14U计算节点正面）；datasheet为给定官方彩页落地页（已HTTP 200校验）；官网原文为组成96或768风冷AI超节点（非8~1024卡），已按官网照录。