AMD MI325X

The AMD Instinct MI325X is a CDNA 3-based accelerator delivering 2.61 PFLOPs (FP8 peak) and 256GB HBM3E memory for leadership AI training and large-scale inference.

The AMD Instinct MI325X is engineered for the most demanding generative AI and HPC workloads: it's a powerhouse accelerator built on the 3rd Gen CDNA 3 architecture. We've equipped this OAM module with an industry-leading 256GB of HBM3E memory and a massive 6 TB/s of peak bandwidth, enabling a single GPU to handle trillion-parameter models. For raw compute, expect up to 2.61 PFLOPs of FP8 peak performance (with structured sparsity, up to 5.22 PFLOPs), all optimized through the open-source ROCm™ software ecosystem. This is the next-gen solution for data centers scaling up large language model (LLM) training and high-throughput inference.

https://www.amd.com/en/products/accelerators/instinct/mi300-series/mi325x.html