At Computex 2024, AMD unveiled an expanded roadmap for its AMD Instinct accelerator family, detailing plans for enhanced AI performance and memory capabilities through annual updates.
“With our updated annual cadence of products, we are relentless in our pace of innovation, providing the leadership capabilities and performance in the AI industry and our customers expect to drive the next evolution of data center AI training and inference,” said Brad McCredie, corporate vice president, Data Center Accelerated Compute, AMD.
2024: The AMD Instinct MI325X accelerator, set to launch in Q4, will feature 288GB of HBM3E memory and 6 terabytes per second of memory bandwidth. This model will continue the use of AMD’s Universal Baseboard server design and is expected to outperform competitors with 1.3 times better compute performance and superior memory capacity and bandwidth.
2025: The AMD Instinct MI350 series will introduce the new AMD CDNA 4 architecture. The MI350X model, built on advanced 3nm process technology, promises up to a 35-fold increase in AI inference performance compared to the current MI300 series. It will also support up to 288GB of HBM3E memory and new AI datatypes.
2026: The roadmap extends to the AMD Instinct MI400 series, powered by the forthcoming AMD CDNA “Next” architecture. These accelerators will bring additional improvements in performance and efficiency, addressing the need for large-scale AI training and inference.
Market adoption and software integration
AMD’s MI300X accelerators are gaining traction with major partners like Microsoft Azure, Meta, and Lenovo. According to McCredie, this adoption reflects the strong performance and value of AMD’s offerings.
The AMD ROCm 6 software stack is playing a key role in this momentum, enhancing the performance of popular large language models (LLMs). For instance, servers equipped with eight MI300X accelerators running ROCm 6 show 1.3 times better inference performance and token generation on Meta’s Llama-3 70B compared to competitors. Additionally, Hugging Face is now testing 700,000 of its models nightly for compatibility with AMD’s accelerators, and AMD continues to support major AI frameworks like PyTorch, TensorFlow, and JAX.