AMD has launched its Instinct MI325X AI accelerator as part of their efforts to compete with Nvidia’s dominance in the AI chip market. The MI325X features 256 GB of HBM3E memory and can achieve up to 2.6 petaFLOPS of FP8 performance, making it ideal for AI workloads, including training large models like Llama 3.1. It was launched in October 2024 and is aimed at improving inference and training tasks for AI, while offering higher memory capacity and bandwidth compared to previous models.
AMD is positioning the MI325X to challenge Nvidia's H200 accelerators, and it is expected to ship in Q4 2024, with broader system availability in early 2025. The company's MI300 series of GPU accelerators has gained significant traction, with AMD raising its full-year outlook for MI300 revenue from $4.0 billion to $4.5 billion.
Comments