AMD Launches New Instinct MI325X Accelerator Card with Advanced Features

TapTechNews October 11th news, the tech media techpowerup released a blog post today (October 11th), reporting that at the AdvancingAI conference, AMD updated and launched the Instinct MI325X accelerator card.

The MI325X accelerator card is based on the CDNA3 architecture and has brought a series of improvements compared to the old MI300X. To meet the trillion-parameter AI model, the AMD MI325X accelerator card focuses on enhancing the HBM3E memory and computing power.

Specifications and Performance

The AMD MI325X accelerator card is equipped with 256GB of HBM3E memory, which is 1.8 times the capacity of MI300 (192GB), and the bandwidth even reaches 6TB/s.

The new accelerator provides 1.3 PetaFLOPS in FP16 and 2.6 PetaFLOPS in FP8 training and inference, which is 1.3 times higher than MI300, all of which are integrated in a chip with 153 billion transistors.

A system equipped with eight MI325X accelerators can achieve 20TB of HBM3E memory and 48TB/s bandwidth, and the computing performance is 10.4 PetaFLOPS (FP16) and 20.8 PetaFLOPS (FP8).

AMD claims that the Instinct MI325X outperforms the NVIDIA H200 HGX system by 1.3 times in memory bandwidth and FP16/FP8 computing performance, and by 1.8 times in memory capacity.

ROCm

The core of the accelerator is the ROCm software stack. AMD plans to introduce ROCm to every GPU (even including consumer GPUs) and cooperate with the open source community to integrate the latest features.

AMD said it will cooperate with the open source community to integrate the features of frameworks such as PyTorch, Triton, and ONNX into the ROCm stack.

Instinct MI350X Series

TapTechNews quoted AMD's official report that the company is preparing the Instinct MI350X series for the second half of 2025, and it is expected to launch the CDNA4 Instinct MI355X accelerator based on TSMC's 3-nanometer process and be equipped with 288GB of HBM3E.