AMD pushes Instinct MI325X with 256 GB HBM3e
Hardware2026-04-15
Source: amd.com
AMD's CDNA-3 refresh widens the memory lead for large-model inference, sharpening the open-ROCm alternative to NVIDIA in the data centre.
AMD's pitch in the data centre has been memory capacity, and the MI325X presses that advantage with 256 GB of HBM3e per accelerator. For inference on very large models, fitting more of the model — and more KV cache — on one device is often worth more than raw FLOPS.
The other half of the story is software: it all runs on the open ROCm / HIP stack, and most CUDA code ports across. See where the parts sit on the AMD prices page.