Accelerator

Instinct MI325X

AMD CDNA 3 oam accelerator summary for training, inference, and roofline-style performance analysis.

Vendor

AMD

Architecture

CDNA 3

Unit

OAM accelerator

Form factor

OAM module

Launch

2024-10-10

Memory

256 GB HBM3E

HBM bandwidth

6 TB/s

BF16 peak

1.3 PFLOPS

FP16 peak

1.3 PFLOPS

FP8 dense peak

2.61 PFLOPS

FP8 sparse peak

5.22 PFLOPS

FP4 dense peak

n/a

FP4 sparse peak

n/a

FP64 peak

81.7 TFLOPS

INT8 peak

2.6 POPS

Interconnect

AMD Infinity Fabric - 8 links, 128 GB/s peak per link

Power

1000 W peak TBP

Software stack

ROCm

Notes

MI325X keeps the CDNA 3 compute profile while increasing HBM capacity and bandwidth relative to MI300X.
Useful comparison point for LLM inference capacity because the memory jump is the headline change.