AMD

Instinct MI325X

Instinct MI325X — illustration of the card's form factor
VRAM
256 GB
FP32 TFLOPS
163.4 TFLOPS
CUDA Cores
19,456
TDP
1000 W

Provider Marketplace

Cheapest
$2.88/hour
Starting from
DigitalOcean logo
DigitalOceanVisit
Best Value
—
Awaiting listings
No additional provider yet
Enterprise Choice
—
Awaiting listings
No additional provider yet

All Cloud Providers

1 Options available
DigitalOcean logo
DigitalOceanCheapest
Reserved · commitment
$2.88/ hour
Estimated Cost
Provision

Estimates only — rates are collected automatically from public provider pages and may be out of date. Prices vary by region, commitment term, and availability, and typically exclude storage, egress, and tax. Confirm current pricing with the provider before purchasing. Last collected .

Compute Performance

FP6481.7 TFLOPS
FP32163.4 TFLOPS
TF32653.7 TFLOPS
FP161307.4 TFLOPS
BF161307.4 TFLOPS
FP82614.9 TFLOPS
INT82614.9 TOPS

Architecture

MicroarchitectureCDNA3
Process NodeTSMC 5nm | 6nm FinFET
Transistors153 Billion
Compute Units304 CUs
CUDA Cores19456
Tensor Cores1216 Matrix Cores
Matrix EngineAMD Matrix Core
Boost Clock2100 MHz
Sparse AccelerationSupported (matrix/structured sparsity)
Dynamic PrecisionSupported (FP8/FP16/BF16/INT8/FP32/FP64)

Memory & VRAM

Memory TypeHBM3E
Total Capacity256 GB
Bandwidth6 TB/s
Bus Width8192-bit
ECC SupportYes (Full-Chip)
Unified MemoryYes (Coherent shared memory on UBB)
Memory PoolingCoherent shared memory (UBB)

Connectivity & Scaling

InterconnectInfinity Fabric
IB Bandwidth128 GB/s
PCIe InterfacePCIe Gen 5 xx16
TopologyRing
Max GPUs/Node8
P2P MemoryCoherent shared memory between accelerators (128 GB/s bidirectional)

Virtualization

SR-IOVSupported (up to 8 partitions)
GPU SharingSR-IOV (up to 8 partitions)

Power & Efficiency

TDP1000 W
Peak Power1000
PSU Required54V UBB
Connectors54V UBB
Thermal LimitsCooling Passive OAM

Physical Design

Form FactorOAM
CoolingPassive

Software Ecosystem

PyTorchSupported
TensorFlowSupported
JAXSupported
HuggingFaceSupported
Compiler StackROCm compilers

Server & Deployment

OEM AvailabilitySeamless, drop-in compatibility with the AMD Instinct MI300X Platform comes through strong support from our OEM partners, industry-leading frameworks, and thousands of large language models.
PreconfiguredThe MI325X is available in an AMD Instinct MI325X Platform with eight accelerators interconnected by an AMD Universal Base Board (UBB 2.0) with HGX host connectors.
DGX/HGXThe MI325X is available in an AMD Instinct MI325X Platform with eight accelerators interconnected by an AMD Universal Base Board (UBB 2.0) with HGX host connectors.

System Compatibility

Required PCIePCIe Gen 5 (x16)
MotherboardOAM module; compatible with AMD Universal Base Board (UBB 2.0) with HGX host connectors
Rack PowerTypical Board Power (TBP) 1000W Peak

Benchmarks & Throughput

Structured Sparsity

Supported (examples in the document show increased peak performance with structured sparsity)

Scaling Efficiency

Multi-GPU connectivity and coherent shared memory described (Infinity Fabric links between GPUs, PCIe Gen5 host interface, 128 GB/s per-link bandwidth)

Multi-GPU Scalability

Scaling Characteristics

Network BottlenecksPCIe Gen5 x16 host interface (128 GB/s); Infinity Fabric link bandwidth 128 GB/s; seven Infinity Fabric links per GPU
ParallelismSR-IOV (up to 8 partitions); coherent shared memory across eight accelerators on a UBB with 128 GB/s bidirectional bandwidth between each GPU; AMD Infinity Fabric links and ROCm support for parallelizing across multiple GPUs/servers

Workload Readiness

LLM Training

Yes

LLM Inference

Yes

HPC / Simulation

Yes

Scientific Computing

Yes

Real-Time Serving

Yes

Market Authority

Key Strengths

Limitations

Also in the Lineup

AMD
Instinct MI100
AMD
Instinct MI200
AMD
Instinct MI210 PCIe Gen4 Passive Accelerator
AMD
Instinct MI250
AMD
Instinct MI250X
AMD
Instinct MI300A APU
AMD
Instinct MI300X
AMD
Instinct MI355X

Expert Insight

The Instinct MI325X represents a powerful alternative for diversified workloads. When comparing cloud providers, consider not just the hourly rate, but also the interconnect bandwidth (InfiniBand/NVLink) and regional availability which can significantly impact total cost of ownership for large-scale training.

Glossary Terms

FP32 TFLOPS
VRAM
TDP
Cores
Information updated daily. Cloud pricing subject to vendor availability.