en.Wedoany.com Reported - U.S. semiconductor design company AMD is strengthening its open-source ecosystem collaboration to address software dependency issues, which are considered a major barrier to entry in the artificial intelligence industry. The company aims to counter proprietary software architectures and increase market share by expanding its open alliance.
At AMD's "Advancing AI 2026" (AAI 2026) conference, to be held in San Francisco, USA, from the 22nd to the 23rd local time, software and network standardization have become another core focus. Recognizing that software compatibility is as important as hardware performance in the infrastructure market, AMD plans to unveil a series of developer-friendly policies and technologies at this event.
One noteworthy aspect is the composition of invited speakers. To demonstrate the versatility of its AI software stack "ROCm 7," AMD has invited influential open-source framework leaders from the industry as speakers, including George Hotz, founder of Tiny Corp, which leads the deep learning framework "tinygrad"; Simon Mo, CEO of Inferact and a major contributor to "vLLM," widely used in global generative AI inference engines; and Ying Sheng, CEO of RadixArc, who previously served on Elon Musk's xAI inference team and created the high-performance inference system "SGLang." They plan to share empirical data on the compatibility and cost-effectiveness of the AMD hardware and ROCm 7 software combination in large language model (LLM) and large-scale training environments.
AMD will also launch a practical project to accelerate the expansion of the developer ecosystem. During the event, a "ROCm Certified Associate" certification course will be offered for the first time to developers attending in person. The core of this program is technical training, helping developers smoothly convert code originally based on NVIDIA's widely used "CUDA" to the AMD ROCm environment. The goal is to allow developers to experience firsthand that the software modification burden of infrastructure migration is not significant, thereby gradually reducing dependence on existing software ecosystems.
AMD will publicly unveil a technology to address network bottleneck issues when connecting tens of thousands of GPUs to build large-scale AI factories: the next-generation Ethernet transport protocol "MRC (Multipath Reliable Connection)." MRC is a technology that operates on a common standard Ethernet architecture, rather than a vendor-specific closed specification. It efficiently distributes data packets and responds to network failures in milliseconds, reducing GPU idle time waiting for data transmission and maximizing overall system efficiency. Some analysts believe that AMD is building a unified platform encompassing server CPUs, networks, and the open-source camp. This conference is expected to declare, through its open alliance, that AMD is ready to serve as a practical alternative in the general-purpose standard infrastructure market.
