The MI325X's 256GB HBM3e — 33% more than MI300X — enables serving 70B+ parameter models with larger KV-caches and batch sizes. The upgraded memory runs at 6 TB/s, making it the ideal drop-in upgrade for existing MI300X infrastructure.
Positioned as a direct competitor to NVIDIA H200, the MI325X delivers comparable FP8 performance with 81% more memory capacity. For inference-heavy deployments, this translates to serving more concurrent users per GPU at lower total cost.
With the same 304 compute units as MI300X and 6 TB/s memory bandwidth, the MI325X accelerates memory-bound HPC workloads. The additional memory capacity enables larger simulation domains without out-of-memory constraints.
Fully compatible with the ROCm software stack and existing MI300X infrastructure, the MI325X is a drop-in upgrade that delivers more memory for organizations invested in the AMD open-source AI ecosystem — training Llama, Mistral, and other open models.
| GPU Architecture | AMD CDNA 3 |
| Process Node | TSMC 5nm / 6nm (3D Chiplet) |
| Compute Units | 304 |
| Stream Processors | 19,456 |
| Matrix Cores | 1,216 (AI Accelerators) |
| Memory Capacity | 256 GB HBM3e |
| Memory Interface | 8192-bit |
| Memory Bandwidth | 6.0 TB/s |
| Infinity Cache | 256 MB |
| Form Factor | OAM (OCP Accelerator Module) |
| Thermal Design Power | 1000W |