IT Home (IT之家) May 19 news - Today, the AMD AI Developer Conference (AMD AI DevDay 2026) officially kicked off in Shanghai. This is the first time this global AI developer conference has been held in China. IT Home (IT之家) was also invited to the venue, and the first impression was the huge popularity - the number of attendees far exceeded expectations.

Specifically, the conference is themed "Open - Build - Innovate - Connect", bringing together industry leaders, technology leaders, ecosystem partners, AMD engineers, and numerous AI developers. Dr. Lisa Su, Chair and CEO of AMD, also attended and delivered the opening keynote speech. Together with industry guests such as Dr. Kai-Fu Lee, founder & CEO of 01.AI (零一万物) and Chairman of Innovation Works (创新工场), she shared the vision for the future development of AI.

This year marks the 20th anniversary of the establishment of AMD's Shanghai R&D Center, and this conference also serves as a platform for AMD to showcase its technology investment and ecosystem development achievements in the Chinese market.

AMD stated that artificial intelligence is currently at a turning point, with the rise of reasoning and agentic AI driving a profound shift in computing demands and prompting a redefinition of the CPU's role. Through its leading product roadmap, AMD is delivering AI performance, collaborating deeply with large AI companies and partners to build future computing power, and leveraging an open platform to achieve broad adaptability, advancing AI development in data centers, embedded systems, edge computing, and intelligent terminals. In the Chinese market, AMD will focus on empowering and cultivating innovative AI developers, rolling out multiple initiatives to build a localized AI ecosystem.

At the core of the conference, Dr. Lisa Su and Dr. Kai-Fu Lee engaged in a dialogue titled "A New Paradigm for AI Agents," exchanging views on key topics such as multi-agent technology development, the AI developer ecosystem, future enterprise organizational structure transformation, and the upgrade of productivity paradigms.

During the conversation, it was noted that with the rise of AI agents, local on-device AI computing is gradually moving toward system-level collaboration, giving birth to a new category in the PC space — the agent host. Such devices require powerful CPU+GPU dual-engine computing power as well as high-bandwidth, large-capacity unified memory to support the operation of complex local large models.

AMD demonstrated the technical advantages of the Ryzen AI Max+ series processors in this field at the conference. The AI agent hosts based on this series have formed a complete product lineup, covering all-in-one PCs, laptops, Mini AI workstations, and other types, fully spanning different usage scenarios and operating system application ecosystems.

Currently, manufacturers including HP, ASUS, Lenovo, Acer, and multiple local innovative brands have launched over 35 related product designs. These devices equipped with the Ryzen AI Max+ series processors support up to 96GB of dedicated GPU memory, featuring low latency and low power consumption. They can natively run local models with up to 200 billion parameters, delivering AI services without relying on the cloud while ensuring user data privacy.

To facilitate the entire path of AI development and deployment, AMD highlighted the latest progress of its open-source software platform ROCm at the conference.

ROCm is a unified software platform that supports all AMD GPUs, enabling interconnectivity from laptops and workstations to data centers. By supporting the HIPCC compiler, ROCm libraries, AI frameworks like PyTorch, and agent frameworks such as OpenClaw, it allows developers to "write once, run across the full path," thereby improving the efficiency of development, testing, and deployment.

At this conference, AMD announced that ROCm now supports the new generation of Ryzen AI 400 series processors, with relevant components available for direct download in ComfyUI.

Starting from ROCm 7.2, the platform has further expanded compatibility with Windows and Linux operating systems. The new PyTorch version can also be easily obtained through AMD software, lowering the deployment barrier for developers in the Windows environment.

In addition, AMD also showcased supporting hardware products, including the Radeon AI PRO R9700 GPU based on the RDNA 4 architecture, equipped with 32GB of VRAM, providing performance support for local AI inference, development, and other memory-intensive workloads. The Ryzen Threadripper PRO 9000 series processors support up to 128 PCIe 5.0 lanes, making them suitable for advanced configurations with multiple GPUs and NVMe storage, capable of meeting the needs of local AI fine-tuning, inference, and application development.

At the same time, AMD released an AI development guide for both Windows and Linux platforms, supporting Ryzen AI and Radeon GPUs, helping developers quickly build AI applications in their existing work environments.

In addition, the conference featured multiple parallel sessions in the afternoon, providing opportunities for in-depth exchange among developers from different directions. Among them, eight GPU hands-on workshops covering three major scenarios—cloud GPUs, edge Radeon GPUs, and on-device Ryzen AI—were all based on the AMD ROCm open-source platform, demonstrating the practical paths of AMD GPUs in various AI development scenarios.

The technical seminar brought together technical lecturers from leading open-source communities, top large model companies, universities, and AMD, sharing insights on key issues in AI infrastructure. The content covers multiple frontier areas such as large model inference optimization, token cost control, AI kernel development, multimodal model optimization, distributed training, MoE training, GPU kernel agents, and multimodal reinforcement learning.

As a unique part of this conference for China, the "Works Speak" developer sub-forum focused on edge-side agent topics, targeting the agent host ecosystem. Through sessions such as practical experience sharing, roadshows by winning teams of the innovation challenge, and a joint practice workshop by AMD and DataWhale, it demonstrated the development potential of edge AI and local agents from three dimensions: experience, works, and hands-on practice.

Advertising Disclaimer: This article contains external jump links (including but not limited to hyperlinks, QR codes, passwords, etc.) for the purpose of delivering more information and saving selection time. The results are for reference only. All articles on IT Home (IT之家) contain this disclaimer.