Moor Threads: Domestic GPU WanCard Cluster is Here

Author: Zhang Yifan

Editor: Shen Siqi

Source: Hard AI

From July 4th to 7th, the 2024 World Artificial Intelligence Conference (WAIC) was held in Shanghai. In addition to AI large models, this year's WAIC also featured a highlight in hardware.

Domestic GPU manufacturer Moore Threads, with the theme "Accelerating a Better World with Full-Stack AI," made its debut at WAIC for the first time with full-stack AI products and the WanKa cluster solution. They showcased the Moore Threads full-stack AI products, including computing acceleration cards, servers, hyper-converged integrated machines, WanKa cluster solutions, and AIGC applications, as well as jointly demonstrated rich industry large models and application solutions based on the QuE AI cluster with many industry partners.

1. Debut of Full-Stack AI Product Line

In just three and a half years since its establishment, Moore Threads has established a full-stack AI computing product line from chips, acceleration cards, servers, clusters to software.

The full-featured GPU chip adopts the advanced MUSA architecture, integrating AI computing acceleration, graphics rendering, video encoding and decoding, physical simulation, and scientific computing.

At this exhibition, the company showcased the following key products:

• Large Model AI Acceleration Card MTT S4000: Specifically designed for large models, supporting 48GB of memory per card;

• AI Large Model Training and Inference Integrated Machine MCCX D800: Dual 8-card GPU server;

• AI Hyper-Converged Integrated Machine KUAE FUSION: A flexible deployment solution integrating inference, training, and fine-tuning;

• QuE (KUAE) WanKa AI Cluster: A large-scale model training platform scalable to WanKa size;

2. Simultaneous Focus on Hardware and Software, from Basic Computing Power to AIGC Applications

Moore Threads' booth is divided into three main sections ——

• QuE Platform: WanKa AI computing base;

• AIGC: Accelerating the release of creative productivity;

• AI+ Industry Digitalization Upgrade;

1) QuE Platform: WanKa AI Computing Base

The "QuE Platform" is a solution that covers the entire AI stack. As an AI computing base, it demonstrates powerful performance and wide compatibility. It includes three core products: QuE cluster management platform, QuE model service platform, and QuE large model inference platform:

• QuE cluster management platform: Achieving automation of intelligent computing center operations;

• QuE model service platform: Providing lifecycle management for large models;

• QuE large model inference platform: Supporting mainstream inference frameworks;

The QuE cluster management platform (KUAE Platform) achieves flexible management of multi-data center, multi-cluster computing resources, integrates multi-dimensional operation monitoring, alarms, and logging systems, and helps the intelligent computing center achieve operational automation.

The QuE model service platform (KUAE ModelStudio) covers the entire process of large model pre-training, fine-tuning, and inference, supporting all mainstream open-source large models, and achieving good compatibility with the CUDA application ecosystem through the MUSIFY code porting tool The Kuaye Big Model Reasoning Platform is based on the efficient MT Transformer engine, supporting industry-leading vLLM reasoning frameworks and self-developed MUSA Serving reasoning frameworks, achieving support for hot technologies such as long-text reasoning, dynamic continuous batch, and MoE.

2) AIGC: Accelerating Creative Productivity

In terms of AIGC creative tools, Moore Threads also showcased products such as content creation and image generation.

"MoBi Ma Liang" is an AIGC content creation platform that integrates self-developed large language models and advanced image generation technology. This platform, based on the Kuaye ZhiSuan cluster as the computing power base, successfully deployed the self-developed MUSAChat large language model, which can complement prompt engineering, polish and translate user input text. The platform can flexibly call upon the capabilities of the SDXL and SD1.5 models to generate detailed images.

The "Creative Portrait" feature can quickly generate personalized portraits. Based on the SDXL model and combined with various IP-Adapter and ControlNet technologies, users only need to upload a photo and describe the desired style to obtain a personalized portrait within 1 minute.

"MoBi Tian Shu" provides a fully automatic picture book generation solution. By inputting a title and a brief story outline, it can generate a complete story, picture book images, narration, subtitles, and background music with one click.

In addition, Moore Threads has also developed the MT AIReality rendering platform, aiming to innovate the asset production process in fields such as film and animation, completing high-quality real-time rendering at a lower cost. It is worth mentioning that Moore Threads has also participated in the Open Sora Plan, using its Kuaye ZhiSuan cluster to provide powerful computing support for AI video generation, aiming to harness the power of the open-source community to reproduce Sora's text-to-video content.

3) AI+ Industry Digitalization Upgrade

In terms of industrial applications, Moore Threads showcased landing solutions in multiple industries such as transportation, finance, and security.

The "Shu Sheng Feng Wu" large model, developed in collaboration with the Shanghai Artificial Intelligence Laboratory, achieved modeling and forecasting of global weather at the 10-kilometer level for over 10 days, and completed a rapid ecological migration from CUDA to MUSA within 24 hours.

In the smart transportation field, the holographic intersection solution developed in collaboration with JiaDu Technology is based on a three-dimensional high-precision map, combined with JiaDu's self-developed ZhiXing large model, achieving real-time traffic information dissemination and intelligent processing.

In the financial services sector, Moore Threads provided efficient and stable large model online services for Reportify, withstanding high traffic business impacts and significantly improving data processing efficiency.

Furthermore, Moore Threads also showcased applications in smart security, AI-assisted decision-making, and other fields. The company's full-function GPU provides diverse computing support in artificial intelligence, video encoding and decoding, meeting the needs of smart security systems for various modal data inputs. The YaYi large model developed in collaboration with Zhongke Wenge demonstrates low latency and high accuracy in areas such as policy interpretation, public opinion perception, government governance, and financial analysis