Atlas 350

The Atlas 350 accelerator card is powered by the Ascend 950PR. Through chip-architecture upgrades it delivers strong AI inference with high compute and large memory, supports SIMT and SIMD programming, and improves the Vector compute ratio. It offers 1.56 PFLOPS mxFP4 for search/recommendation and multimodal generation inference.

Key Highlights

• Strong compute: industry-leading 1561 TFLOPS @mxFP4
• SIMT support improves logic/memory-disjoint operator performance and developer ease
• HiF8/mxFP8/mxFP4 formats boost quantization. High bandwidth: 1.4TB/s memory bandwidth, PCIe 5.0, and up to 424GB/s card-to-card Lingqu interconnect.

Technical Specifications

Typical Use Cases

英文站hiascend.com无Atlas 350独立英文产品页（/en路径均重定向en-sg首页），英文文案为翻译稿；官网无独立PDF彩页/技术白皮书链接；img1取自官网350页视角2斜侧渲染图（视角1正面图官网与Atlas 150共用同一静态文件img-Atlas15001，故选用唯一冠名350的img-Atlas35002）。