Documentation
中文

Challenges

High Computing Costs
With over 10 million cards expected, compute capacity is expanding to meet growing training needs and surging inference demand.
Rapid Application Innovation
AI drives rapid innovation in Internet foundation models and applications, with frequent releases in mainstream ToC scenarios—some already commercialized.
High Performance Required
Internet services—including intelligent assistants, search, and multimodal applications—depend on low-latency, high-throughput AI platforms.

Solutions

LLM
Ultimate experience of fast, easy, and accurate training and inference
Multimodal
High performance in understanding and generation scenarios
SRA
High availability, strong generalization, and ease of use

Related Products

Atlas 800T A2 Training Server
Learn More
Atlas 800I A2 Inference Server
Learn More
Atlas 300I Duo Inference Card
Learn More

Application Scenarios

Intelligent Assistant
Supports 100B-parameter model inference, with millions of active users and hundreds of millions of daily requests.
Supports high-performance and low-latency dialog interactions.
SRA
Enhances generalization to boost performance in typical recommendation models.
Speeds up innovation in generative recommendation for the Internet industry.
Smart Office
Delivers a fast, user-friendly inference solution for document typesetting, copywriting, slide creation, and formula layout.
Text-to-Image Generation
Generates images in seconds, streamlining poster design and avatar creation.
Delivers strong performance for fast text generation.
Lowers staffing costs for design and customer support teams.

Use Cases

Deep Collaboration Between Ascend and ERNIE Series Models
On June 30, 2025, the ERNIE 4.5 series was open-sourced with technical support from Ascend. Following the release, Ascend AI enabled post-training for ERNIE-4.5-300B-A47B, ensuring seamless integration with Ascend.
Ascend × 360 Group: Nano AI to Accelerate Multimodal Search
Powered by high-performance hardware, fused operators, MindIE acceleration algorithms, and DiT quantization, the Ascend inference solution enables 360 Group to build China's first super search agent—Nano AI. It enhances core multimodal models like text-to-image, text-to-video, and multimodal understanding, delivering twice the industry-standard performance.
View details
Ascend × WPS: High-Efficiency Smart Platform for Office Tasks
Built on Ascend, WPS AI pre-trained its document writing model from scratch, achieving stable monthly training. Within days, it deployed full-link inference across PowerPoint, Word, and Excel—enabling fast document typesetting, copywriting, slide creation, and formula layout to simplify office work.
Ascend AI × LinkSure: Integrating Intelligence into Diverse Life Scenarios
Ascend provides a MoE EP solution with 320 tokens/s single-card throughput and 50–100 ms latency, supporting 128K ultra-long sequences. This enables LinkSure to embed AI into everyday scenarios like messaging, bedtime stories, FaceFun, and GEAK smart cloud.