| Product SKU | MI325X-256GB |
|---|---|
| Name | AMD Instinct™ MI325X |
| Family | Instinct |
| Series | Instinct MI300 Series |
| Form Factor | Servers |
| Launch Date | 10/10/2024 |
| GPU Architecture | CDNA3 |
| Lithography | TSMC 5nm, 6nm FinFET |
| Stream Processors | 19,456 |
| Matrix Cores | 1216 |
| Compute Units | 304 |
| Peak Engine Clock | 2100 MHz |
| Peak Eight-bit Precision (FP8) Performance (E5M2, E4M3) | 2.61 PFLOPs |
| Peak Eight-bit Precision (FP8) Performance with Structured Sparsity (E5M2, E4M3) | 5.22 PFLOPs |
| Peak Half Precision (FP16) Performance | 1.3 PFLOPs |
| Peak Half Precision (FP16) Performance with Structured Sparsity | 2.61 PFLOPs |
| Peak Single Precision (TF32 Matrix) Performance | 653.7 TFLOPs |
| Peak Single Precision (TF32) Performance with Structured Sparsity | 1.3 PFLOPs |
| Peak Single Precision Matrix (FP32) Performance | 163.4 TFLOPs |
| Peak Single Precision (FP32) Performance | 163.4 TFLOPs |
| Peak Double Precision Matrix (FP64) Performance | 163.4 TFLOPs |
| Peak Double Precision (FP64) Performance | 81.7 TFLOPs |
| Peak INT8 Performance | 2.6 POPs |
| Peak INT8 Performance with Structured Sparsity | 5.22 POPs |
| Peak bfloat16 | 1.3 PFLOPs |
| Peak bfloat16 with Strutured Sparsity | 2.61 PFLOPs |
| Transistor Count | 153 Billion |
| External Power Connectors | 54V UBB |
| Typical Board Power (TBP) | 1000W Peak |
| Last Level Cache (LLC) | 256 MB |
| Dedicated Memory Size | 256 GB |
| Dedicated Memory Type | HBM3E |
| Infinity Cache | Yes |
| Memory Interface | 8192-bit |
| Memory Clock | 6 GHz |
| Peak Memory Bandwidth | 6 TB/s |
| Memory ECC Support | Yes (Full-Chip) |
| GPU Form Factor | OAM Module |
| Bus Type | PCIe® 5.0 x16 |
| Infinity Fabric™ Links | 8 |
| Peak Infinity Fabric™ Link Bandwidth | 128 GB/s |
| Cooling | Passive OAM |
| Supported Technologies | AMD CDNA™ 3 Architecture , AMD ROCm™ – Ecosystem without Borders , AMD Infinity Architecture |
| RAS Support | Yes |
| Page Retirement | Yes |
| Page Avoidance | Yes |
| SR-IOV | Yes |
GPU
AMD Instinct MI325X 256GB
The AMD Instinct MI325X 256GB is a data-center accelerator built on the CDNA3 architecture, with 256 GB HBM3E of memory and 6 TB/s of memory bandwidth. It combines 2.61 PFLOPs FP8 and 1.3 PFLOPs FP16 Tensor Core performance with a board power of 1000 W peak in a Servers form factor.
Overview
The AMD Instinct MI325X 256GB is a data-center accelerator built on the AMD CDNA 3 architecture with 256 GB of HBM3E memory and a peak memory bandwidth of 6 TB/s. It pairs 304 compute units and 1,216 matrix cores with 2.61 PFLOPs of FP8 and 1.3 PFLOPs of FP16 performance (5.22 PFLOPs and 2.61 PFLOPs respectively with structured sparsity) in a 1,000 W OAM module for large-scale AI training and inference.
At a glance
- Architecture: AMD CDNA 3; series: Instinct MI300 Series
- Memory: 256 GB HBM3E, 6 TB/s peak bandwidth, 8192-bit interface, full-chip ECC; 256 MB last-level cache
- Compute: 304 compute units, 19,456 stream processors, 1,216 matrix cores
- Peak performance: FP8 2.61 PFLOPs (5.22 PFLOPs sparse), FP16/BF16 1.3 PFLOPs (2.61 sparse), FP64 81.7 TFLOPs
- Form factor: passive OAM module; board power 1,000 W peak; 54 V UBB power
- Bus: PCIe 5.0 x16; 8x Infinity Fabric links (128 GB/s each)
- Software: AMD ROCm; PCIe SR-IOV supported; RAS with page retirement and page avoidance
When to choose the MI325X
Choose the MI325X when the priority is the largest possible memory footprint per accelerator for very large models — 256 GB of HBM3E with 6 TB/s — within an open ROCm software stack and an OAM-based server design. The 8 Infinity Fabric links make it suited to tightly coupled multi-GPU nodes for training and high-throughput inference.
Platform & compatibility
The MI325X is a passive OAM module powered over 54 V UBB, so it is integrated into OEM accelerator servers rather than installed as a PCIe card (the PCIe 5.0 x16 bus is the host link). It is programmable through AMD ROCm and supports PCIe SR-IOV along with RAS features including page retirement and page avoidance.
FAQ
How much memory does it have? 256 GB of HBM3E with 6 TB/s of bandwidth and full-chip ECC.
Is it a PCIe card? It is a passive OAM module integrated into OEM servers, with a PCIe 5.0 x16 host bus.
How is it programmed? Through the AMD ROCm software ecosystem.
Related
- All AMD Instinct accelerators
- Data-center GPU buying guide: A100 to B300
- Browse all GPUs and accelerators
Configuration, availability and lead time depend on the required quantity, server pairing and delivery destination. Send your requirement to receive a quotation.
