AMD · GPUs / AI accelerators
AMD Instinct MI325X
Data-centre accelerator; enquire with a compatible system platform
Accepting enquiries · sourcing on request
2 official PDFsManufacturer images
GET TO KNOW THIS PRODUCT
Product overview
AMD Instinct MI325X is a CDNA 3 OAM accelerator for data-centre AI and HPC with 256 GB of HBM3E memory.
At a glance
- GPU architectureAMD CDNA 3
- GPU memory256 GB HBM3E ECC
- Memory bandwidth6 TB/s
AMD also documents an eight-MI325X UBB 2.0 platform with 2.048 TB total HBM3E. The specifications below describe one module; complete server configurations are confirmed separately.
Typical applications
Data-centre AI training, inference and HPC; select with ROCm and compatible system requirements.
Plan your configuration
Share your model or application, host system and workload so we can check compatibility and configuration.
Technical specifications
Values are sourced from AMD product pages and MI325X datasheets. English technical field names follow the manufacturer; Chinese fields use AMD Traditional Chinese terminology with translated PDF additions.
GPU compute7 fields
- GPU Architecture
- AMD CDNA 3
- Lithography
- TSMC 5 nm / 6 nm FinFET
- Compute Units
- 304
- Matrix Cores
- 1,216
- Stream Processors
- 19,456
- Peak Engine Clock
- 2,100 MHz
- Transistor Count
- 153 billion
Source: AMD product specifications
Memory and cache6 fields
- Dedicated Memory
- 256 GB HBM3E
- Memory Interface
- 8,192-bit
- Peak Memory Bandwidth
- 6 TB/s (theoretical peak)
- Memory Clock (datasheet)
- Up to 6.0 GT/s
- Last Level Cache / Infinity Cache
- 256 MB
- Memory ECC Support
- Full-chip ECC
Interfaces, power and cooling6 fields
- GPU Form Factor
- OAM Module
- Bus Type
- PCIe 5.0 x16
- Scale-up AMD Infinity Fabric Links (datasheet)
- 7 × 128 GB/s; bidirectional bandwidth per link
- Maximum TBP
- 1,000 W (peak)
- External Power Connectors
- 54V UBB
- Cooling
- Passive OAM
Peak theoretical compute performance9 fields
Single-MI325X theoretical peaks from AMD. Sparse figures require the corresponding structured-sparsity operation; these are not measured application results.
- TF32 matrix
- Dense: 653.7 TFLOPs Structured sparsity: 1,307.4 TFLOPs
- FP16
- Dense: 1,307.4 TFLOPs Structured sparsity: 2,614.9 TFLOPs
- BFLOAT16
- Dense: 1,307.4 TFLOPs Structured sparsity: 2,614.9 TFLOPs
- FP8
- Dense: 2,614.9 TFLOPs Structured sparsity: 5,229.8 TFLOPs
- INT8
- Dense: 2,614.9 TOPS Structured sparsity: 5,229.8 TOPS
- FP64 vector
- 81.7 TFLOPs
- FP32 vector
- 163.4 TFLOPs
- FP64 matrix
- 163.4 TFLOPs
- FP32 matrix
- 163.4 TFLOPs
Source: AMD datasheet (pp. 1–2)
Reliability, virtualization and software5 fields
Video decoding requires compatible media software. Check operating-system, driver and framework versions against the selected ROCm release.
- RAS features
- ECC, Page Retirement, Page Avoidance
- Virtualization support
- SR-IOV; up to 8 partitions
- Decoders
- 4 groups; HEVC/H.265, AVC/H.264, VP9 or AV1
- JPEG/MJPEG CODEC
- 32 cores; 8 cores per group
- Supported Technologies
- AMD ROCm
Additional information checked:
MANUFACTURER DOCUMENTS
Official PDFs & product information
Documents open on the manufacturer’s site in a new tab, in their original language. Family documents may include options not supplied in every configuration.
Sources & review dates
Specification source reviewed: 2026-10-21 · Product page published: 2026-10-23
Additional PDF links checked: 2026-10-26
Additional official PDF · www.amd.comAdditional official PDF · www.amd.comAdditional manufacturer reference · www.amd.comPLAN YOUR PURCHASE
Confirm the configuration before ordering
Tell us the quantity, intended use, delivery date and location. We will verify the SKU, available configuration, lead time, regional supply and warranty before quoting.
Submitting an enquiry is not an order, reservation or payment.
You may also like
Explore similar products, with photos and full specifications.
