AMD adds ultra-fast memory to flagship AI Instinct accelerator as it looks to next-generation CDNA 4 architecture – Instinct MI325X accelerator has 2x memory and 30% more bandwidth compared to Nvidia’s H200
AMD has unveiled new CPU, NPU and GPU architectures aimed at “powering end-to-end AI infrastructure from the data center to PCs,” alongside an expanded AMD Instinct accelerator roadmap and a new Instinct MI325X accelerator, which the company says will be available in Q4 2024.
The new Instinct MI325X offers 288 GB of HBM3E memory and 6 TB/s of memory bandwidth. AMD says this means it will offer 2x the memory capacity and 1.3x bandwidth than “the competition,” by which it means Nvidia’s H200, as well as 1.3x better compute performance.
The memory upgrade is the main change here as it will use the original CDNA 3 architecture like the MI300X and the clock speeds also appear to be unchanged at 2.1GHz.
Looking to the future
After the Instinct MI325X comes the Instinct MI350 series. Expected to be available in 2025, it will be powered by the new CDNA 4 architecture, which AMD says will deliver up to 35x higher AI inference performance compared to the Instinct MI300 series.
That will be followed in 2026 by the AMD Instinct MI400 series which will be based on AMD’s CDNA Next-Gen architecture. The company understandably didn’t go too deep into this.
“The AMD Instinct MI300X accelerators continue their strong adoption by numerous partners and customers, including Microsoft Azure, Meta, Dell Technologies, HPE, Lenovo and others, a direct result of the exceptional performance and value proposition of the AMD Instinct MI300X accelerator” , said Brad McCredie, corporate vice president, Data Center Accelerated Compute, AMD.
“With our updated annual product offering, we are relentless in our pace of innovation, delivering the leadership capabilities and performance that the AI industry and our customers expect to drive the next evolution of AI training and inference in data centers.”