MI325X GPU

The AMD Instinct MI325X is a CDNA 3-based data-center accelerator built on a 5 nm process, featuring 256 GB of HBM3E memory (6 TB/s peak bandwidth), 304 compute units (19,456 stream processors), and 1,216 matrix cores. It supports PCIe Gen 5 and AMD Infinity Fabric for coherent multi-GPU scaling, and is optimized for large-model AI training/inference as well as HPC workloads.

Starting Price
$2.25/hr
Available on 1 cloud providers
MI325X GPU

Key Specifications

๐Ÿ’พMemory

256GB VRAM

๐Ÿ—๏ธArchitecture

CDNA 3

โš™๏ธCompute Units

304

๐ŸงฎTensor Cores

1216

Technical Specifications

Hardware Details

ManufacturerAMD
ArchitectureCDNA 3
CUDA CoresN/A
Tensor Cores1216
RT CoresN/A
Compute Units304
GenerationGen 3

Memory & Performance

VRAM256GB
Memory Interface8192-bit
Memory Bandwidth6000 GB/s
FP32 PerformanceN/A
FP16 Performance1300 TFLOPS
INT8 PerformanceN/A

Performance

Computing Power

Tensor Cores1,216

Computational Performance

FP16 (TFLOPS)1,300

Benchmark Resources

Pros and Cons

Pros

  • โœ“Very high memory capacity (256 GB HBM3E) and bandwidth (6 TB/s)
  • โœ“High FP16/FP8 throughput
  • โœ“Optimized for large-model AI workloads and HPC

Cons

  • โœ—Extremely high power consumption (up to 1000 W) requiring specialized cooling/power infrastructure
  • โœ—Very high acquisition cost; limited to data-center/enterprise deployments

Key Features

โ€ข256 GB HBM3E memory on an 8192-bit bus
โ€ข6 TB/s peak memory bandwidth
โ€ข304 compute units (19,456 stream processors) + 1,216 matrix cores
โ€ขBuilt on 5 nm process with up to 2.10 GHz clocks
โ€ข256 MB Infinity Cache and PCIe Gen 5 x16 with Infinity Fabric
โ€ขSupports OAM form factor for multi-GPU platforms

About MI325X

The MI325X leverages a dual-chiplet CDNA 3 design to deliver 19,456 stream processors and 1,216 matrix cores clocked up to 2.10 GHz, backed by 256 GB of HBM3E at 6 TB/s. Built on TSMCโ€™s 5 nm node, it supports PCIe Gen 5 ร—16 and AMD Infinity Fabric for coherent multi-GPU scaling. It is optimized for large AI model training and inference, offering native matrix-sparsity support and a shared 256 MB Infinity Cache to reduce latency. This accelerator fits into OAM (Open Compute Accelerator Module) form factors and targets data centers running HPC, ML/DL, and large-language models.

Common Use Cases

High-performance AI/ML training and inference (large-language models, generative AI), HPC workloads, data-center acceleration

Machine Learning & AI

  • Training large language models and transformers
  • Computer vision and image processing
  • Deep learning model development
  • High-performance inference workloads

Graphics & Compute

  • 3D rendering and visualization
  • Scientific simulations
  • Data center graphics virtualization
  • High-performance computing (HPC)

Market Context

The MI325X sits within AMD's CDNA 3 architecture lineup,. It's designed specifically for data center and enterprise use.

Cloud Availability

Available across 1 cloud providers with prices ranging from $2.25/hr. Pricing and availability may vary by region and provider.

Market Position

Released in 2024, this GPU is positioned for enterprise and data center workloads.

Current Pricing

ProviderHourly PriceSource
TensorWave
$2.25/hr

Prices are updated regularly. Last updated: 6/17/2025