On May 19, AMD AI Developer Day was successfully held at the Shangri-La, Qiantan, Shanghai. The event focused on technical exchanges across the entire AI development stack, bringing together leading voices from academia and industry. Wang Yu, Professor of the Department of Electronic Engineering at Tsinghua University, Founder of Infinigence, and IEEE Fellow, participated alongside Dr. Lisa Su, Chair and CEO of AMD; Pan Xiaoming, Senior Vice President of AMD and President of AMD Greater China; Jack Huynh, Senior Vice President and General Manager of Computing and Graphics at AMD; Dr. Kai-Fu Lee, Founder and CEO of 01.AI and Chairman of Sinovation Ventures; and Dr. Yibo Zhu, Co-Founder and CTO of StepFun. Together with AI researchers, developers, and AMD engineers, they engaged in in-depth discussions on how advances in foundation model technologies are driving industry transformation, accelerating innovation, and shaping the future of computing across cloud, edge, and device environments.
Infinigence and AMD are jointly advancing the development of an open technology ecosystem, achieving a series of phased technical milestones and breakthroughs across the cloud, edge, and device computing continuum. On the cloud side, the two parties are co-developing a reinforcement learning infrastructure foundation to support algorithm development for embodied intelligence applications. At the edge layer, they enable efficient model execution across diverse AMD hardware platforms through automated operator generation and inference acceleration techniques. On the device side, the collaboration focuses on local data anonymization, edge–cloud compute coordination, and intelligent model orchestration, accelerating the deployment of AI in real-world scenarios such as AI boxes and embodied robotics. At the event, Zeng Shulin, General Manager of Infinigence Shanghai and Head of the Edge and Device Division, presented the collaboration achievements to Dr. Lisa Su.
01
Jointly Building a Reinforcement Learning Technology Foundation on the Cloud Side for End-to-End Simulation, Training, and Inference
To enable robots to acquire human-like capabilities such as locomotion and manipulation, developers require a powerful “virtual training ground” for large-scale learning and iteration. To this end, Infinigence has deeply integrated AMD compute resources with the RLinf reinforcement learning framework to build an end-to-end platform for embodied intelligence that unifies simulation, training, and inference. The platform is designed to generate core policy models for robotic action and decision-making. RLinf introduces a unified hardware abstraction layer that enables tight integration with AMD GPU architectures, supporting efficient execution of embodied reinforcement learning workloads across AMD Instinct™ MI series and Radeon™ GPU platforms.
The platform has already completed validation of embodied reinforcement learning workflows based on the LIBERO simulation environment and the OpenVLA-OFT model. It provides a fully closed-loop pipeline spanning simulation, reinforcement learning training, and model inference deployment, delivering a stable and high-performance infrastructure foundation for embodied intelligence research and development.
RLinf is a large-scale reinforcement learning framework for embodied intelligence jointly developed by Infinigence, Tsinghua University, and Zhongguancun Academy. It has been widely adopted by leading research groups across both academia and industry in the embodied AI space, and has emerged as one of the most widely used frameworks in the field.
By introducing M2Flow, a novel reinforcement learning programming paradigm based on macro-to-micro flow transformation, RLinf decouples high-level programmable workflows from low-level system optimizations, achieving a balance between flexibility and execution efficiency.
In its v0.1 release, RLinf established an end-to-end pipeline spanning simulation, training, and inference, with unified support for mainstream embodied AI foundation models, reinforcement learning algorithms, and simulation environments. Through system-level optimization, it delivers significant improvements in GPU utilization and overall training throughput, achieving up to a 1.27× gain in benchmark evaluations. In its v0.2 evolution toward real-world deployment, RLinf extends its capabilities from simulation to physical environments, enabling fully heterogeneous and fully asynchronous real-robot reinforcement learning. With an adaptive communication layer and a unified hardware abstraction interface, the system allows robots to be operated with the same flexibility as GPUs, while supporting cross-domain, multi-robot deployment. In addition, RLinf introduces asynchronous training–inference decoupling and human-in-the-loop capabilities, providing system-level infrastructure for scalable real-world reinforcement learning.
02
Operator Generation and Software-Hardware Co-Optimization Across Cloud, Edge, and Devices for Efficient AI Model Inference
Algorithms trained in the cloud must be efficiently adapted to a diverse range of edge and device hardware in order to achieve high-performance inference. However, differences in chip architectures often lead to suboptimal deployment efficiency. To address this challenge, agent-based automatic compilation plays a critical role in bridging the gap between model training and heterogeneous hardware execution. Infinigence integrates AMD’s open-source ROCm software stack with its in-house KernelMind intelligent framework to enable automated code translation and optimization. Built on the Mizar inference acceleration engine, the system can automatically generate optimized deployment code for cloud-trained models across a wide range of AMD hardware platforms, including the AMD Instinct™ MI series, Radeon™ GPUs, and Ryzen™ AI processors. This enables high-performance, energy-efficient inference across cloud, edge, and device environments.
KernelMind is an AI-agent-driven, integrated cloud-edge-device inference acceleration and optimization platform.In engineering practice, KernelMind employs a multi-agent collaboration architecture. A “strategy” agent is responsible for analyzing performance bottlenecks and formulating optimization plans; a “code” agent generates and refactors ROCm/HIP kernel implementations; and a “verification” agent performs compilation, execution, correctness validation, and performance benchmarking on real hardware. Through this closed-loop workflow—“think, execute, verify, and refine”—KernelMind transforms operator optimization from a process dependent on expert intuition into an automated system capable of continuous iteration and knowledge accumulation.
At a deeper system level, the Infinigence Mizar inference acceleration engine integrates parallel decoding, graph-computation fusion optimization, heterogeneous memory scheduling, and KernelMind’s operator-level optimization capabilities. Together, they improve inference efficiency across cloud, edge, and device environments while reducing power consumption and memory overhead.
On real AMD Instinct™ MI-series hardware, KernelMind has been evaluated across key inference operator workloads, including GQA Attention, MLA Attention, GDN, and GEMM. It delivers performance improvements ranging from approximately 1.5× to 2.8× across multiple industrial-grade operator scenarios, demonstrating strong engineering effectiveness within the AMD ecosystem.
03
Building an Edge–Cloud Collaborative Closed-Loop System to Bring AI into the Physical World
On the device side, real-world AI applications require both robust perception and responsive decision-making in dynamic environments, while also meeting strict data privacy and security requirements.Leveraging the heterogeneous compute architecture and high-bandwidth unified memory capabilities of the AMD Ryzen™ AI Max+ 395 platform, Infinigence has built a local inference system for embodied intelligence based on the InfiniClaw edge–cloud collaboration architecture.
Through key technologies such as multimodal model compilation and optimization, heterogeneous operator fusion, dynamic model routing, and edge–cloud coordinated scheduling, the system enables on-device privacy-preserving processing, local multimodal inference, and cloud-assisted generative capabilities. It supports efficient inference and collaborative execution across a range of embodied AI models, including VLMs (Vision-Language Models), VLN (Vision-Language Navigation), and VLA (Vision-Language-Action models), accelerating the deployment of AI in real-world scenarios such as AI boxes and embodied robotics.
Live presentation by Zeng Shulin
In local multimodal privacy protection, Infinigence has pioneered a three-stage edge–cloud secure inference framework—“local anonymization, cloud processing, and local reconstruction”—designed to ensure that sensitive data never leaves the device. User-input multimodal data is first processed on-device, where a local multimodal foundation model and a security knowledge base automatically detect and anonymize sensitive information, producing a fully de-identified request. This request is then securely transmitted to the cloud, where large language models such as GLM-5 and Kimi-K2.5 are invoked for inference and generation, without any exposure to the original private data. Finally, a local reconstruction engine restores the anonymized elements with high fidelity, ensuring semantic consistency while achieving true zero data egress. This design enables strong privacy guarantees while preserving full model capability, achieving a balance of “no data leaves the device, no performance is compromised.”
For intelligent model routing, InfiniClaw is tightly integrated with the Infinigence platform to enable cost-efficient invocation of cloud-based large language models such as GLM-5 and Kimi-K2.5.Through capabilities including task decomposition, global orchestration, complex decision routing, and unified output validation, the system dynamically allocates workloads between local and cloud models. This enables intelligent routing across large and small models, significantly improving efficiency while reducing token usage costs by up to 70%.
04
Jointly Building a Developer Ecosystem to Empower a New Era of “AI Productivity”
In collaboration with AMD, Infinigence has established a full end-to-end development loop spanning cloud-based simulation, training, and inference; automated operator generation and optimization at the edge; and real-time inference on device. Looking ahead, Infinigence plans to progressively open-source its full cloud–edge–device development toolchain, while also developing a dedicated embodied intelligence developer toolkit tailored for the AMD ecosystem.
Jack Huynh (left) in conversation with Professor Wang Yu (right)
Wang Yu, Professor in the Department of Electronic Engineering, Tsinghua University, founder of Infinigence, and IEEE Fellow, stated: “The essence of open source lies in building a win-win ecosystem. The philosophy embraced by my team aligns closely with AMD’s open ROCm vision, which is dedicated to empowering all participants in the ecosystem. Together with AMD, we are building open infrastructure for the next generation of intelligent systems.
Looking ahead, Infinigence will continue to collaborate with AMD to co-develop an open technology ecosystem, with the goal of providing developers with a complete toolchain for “self-evolving agents” and integrated cloud–edge–device intelligence.
As a key strategic partner of AMD, Infinigence will further leverage complementary strengths to advance joint R&D, expand the open ecosystem, and streamline the development and deployment of large-scale AI applications. These efforts aim to deliver stronger differentiated advantages for enterprises and developers, and accelerate the evolution of the global AI ecosystem.
Looking ahead, Infinigence will continue to collaborate with AMD to co-develop an open technology ecosystem, with the goal of providing developers with a complete toolchain for "self-evolving agents" and integrated cloud-edge-device intelligence. As a key strategic partner of AMD, Infinigence will further leverage complementary strengths to advance joint R&D, expand the open ecosystem, and streamline the development and deployment of large-scale AI applications. These efforts aim to deliver stronger differentiated advantages for enterprises and developers, and accelerate the evolution of the global AI ecosystem.
