AMD Scales GLM-5.1-FP8 on 64 MI300X GPUs; UALink Linux Kernel Patches Posted

AMD's ROCm team detailed the successful scaling of the GLM-5.1-FP8 Mixture-of-Experts (MoE) model across 64 AMD Instinct MI300X GPUs. This achievement demonstrates the capabilities of AMD's hardware and ROCm software stack in handling frontier-level sparse AI models at scale. The blog post highlights the critical systems engineering required to serve such large models efficiently, particularly when they exceed the capacity of a single node. The GLM-5.1 is characterized as a large, sparse MoE model designed for long context windows, presenting significant challenges for memory bandwidth and inter-GPU communication. The utilization of FP8 precision was key to managing memory footprints and computational efficiency across the distributed system. This work is a direct response to the increasing demand for high-performance, scalable solutions for large language models and other sparse AI architectures, positioning AMD Instinct MI300X GPUs as a viable platform for demanding AI workloads.

AMD has submitted a substantial patch series to the Linux kernel mailing list, aimed at enabling support for its UALink interconnect technology. This second series of patches indicates significant progress in integrating UALink into the mainline Linux kernel, targeting the upcoming Linux 7.3 merge window. UALink is AMD's answer to high-bandwidth, low-latency interconnects, crucial for advanced computing systems where seamless communication between CPUs, GPUs, and other accelerators is paramount. UALink is expected to play a vital role in future AMD data center and high-performance computing platforms, facilitating faster data exchange and improved overall system performance for complex workloads, including AI training and scientific simulations. The introduction of these kernel patches signifies that the underlying software infrastructure for UALink is maturing, paving the way for wider adoption and hardware utilization in Linux environments. This development ensures that the operating system can effectively manage and leverage the capabilities of UALink for optimal performance.

NVIDIA announced advancements with its Cosmos 3 Edge system, specifically tailored for on-device robot control. This system aims to provide policies that enable robots to dynamically adapt to their sensors, environments, and tasks while running efficiently on onboard computing hardware. The core concept behind Cosmos 3 Edge leverages world models, which offer a foundational approach for robots to understand and interact with their surroundings more intelligently and adaptably. This initiative underscores NVIDIA's continued focus on edge AI and robotics, bridging the gap between sophisticated AI models and real-world autonomous applications. The ability to deploy complex, adaptive control policies directly on device is crucial for robots operating in dynamic and unstructured environments, requiring high-performance, low-latency processing at the edge. By integrating world models, Cosmos 3 Edge seeks to empower robots with enhanced autonomy, allowing them to learn and execute tasks more effectively without constant external intervention, leveraging NVIDIA's GPU technology for local processing.