Released a kilocard-scale heterogeneous chip hybrid training platform, Wuwen Core Dome aims to create the most cost-effective AI infrastructure | WAIC 2024

What role does infrastructure play in the AI ​​era? Some say it's like electricity, some say it's like water.

Wuwen Xinqiong believes that excellent infrastructure is a kind of "magic" that can effectively reduce the cost of large models and allow more people to embrace new technologies.

On July 4, Xia Lixue, co-founder and CEO of Wuwen Core Dome, released the world's first kilo-calorie heterogeneous chip hybrid training platform. The computing power utilization of the kilo-calorie heterogeneous hybrid training cluster reached a maximum of 97.6%.

Four months ago, Wuwen Core Dome’s Infini-AI large model development and service cloud platform announced its first public beta. Customers from large model companies such as Zhipu AI, Dark Side of the Moon, and Shengshu Technology have been stably using Infini-AI. In terms of computing power, there are more than 20 AI Native application startups that continue to call various preset model APIs on Infini-AI and use the tool chain provided by Wuwen Core Dome to develop their own business models.

Building AI Native infrastructure in the era of large models can not only provide AI developers with a more versatile, efficient, and convenient R&D environment, but is also a key cornerstone for achieving effective integration of computing resources and supporting the sustainable development of the AI ​​industry. Compared with the "relatively concentrated" pattern of model layers and chip layers in the world, China's model layers and chip layers are more diverse.

However, diversity also means challenges. A large number of heterogeneous chips have also formed "ecological silos". Different hardware ecosystems are closed and incompatible with each other, which brings a series of technical challenges to users of computing power. This is the biggest difficulty in building AI Native infrastructure, and it is also an important reason why the current large model industry is facing a "computing power shortage".

Wuwen Core Qiong has top-notch AI computing optimization capabilities and computing power solution capabilities, as well as forward-looking judgment on the "M types of models" and "N types of chips" industry patterns, and has taken the lead in building an "MxN" middle layer ecological pattern to achieve Efficient and unified deployment of multiple large model algorithms on multiple chips.

Up to now, Infini-AI has supported more than 30 models such as Qwen2, GLM4, Llama3, Gemma, Yi, Baichuan2, ChatGLM3 series, etc., as well as AMD, Huawei Shengteng, Biren, Cambrian, Suiyuan, Haiguang, Tianshu There are more than 10 kinds of computing cards including Zhixin, Muxi, Moore Thread, and NVIDIA.

"There is no contradiction between pushing up the technical ceiling and the diffusion of technology, and it depends on how we are determined to treat this technology." Xia Lixue said, "When we use various AI applications in the future, we will not know which base models it calls and which models are used. Which kind of accelerator card has the computing power – this is the best AI Native infrastructure."

# Welcome to follow the official WeChat public account of aifaner: aifaner (WeChat ID: ifanr). More exciting content will be provided to you as soon as possible.

Ai Faner | Original link · View comments · Sina Weibo