AMD Instinct MI325X Full Specs

Summary

19,456 Shaders 2.1GHz

256GB HBM3e6.14TB/s

Peak·FP32·FP16 163.4 TFLOPS

Form Factor OAM Module

TDP 1000W

Highlights

Peak OPS

Peak OPS FP32 FP16 Tensor FP8-32 FP8-32 Tensor Sparse Tensor FP16-32 FP16-32 Tensor Sparse Tensor FP32 FP64 Tensor FP64 Tensor BF16 BF16 Tensor Sparse Tensor TF32 Tensor INT8 Memory Bandwidth

Instinct MI325X5.23 POPSINT8 Tensor Sparse

x1

Top Score3.35 PFLOPS FP4 Tensor Sparse

x0.64

Desktop1xGeForce RTX 5090

Theoretical Performance

Clock Speed2100MHz UC·Base·Boost·OC

Peak OPS 5.23 POPSINT8 Tensor Sparse

Tensor FP8-32 2.62 PFLOPS FP8-32 Tensor Sparse 5.23 PFLOPS

FP16 163.4 TFLOPS Tensor FP16-32 1.31 PFLOPS FP16-32 Tensor Sparse 2.62 PFLOPS

Tensor BF16 1.31 PFLOPS BF16 Tensor Sparse 2.62 PFLOPS

Tensor TF32 653.7 TFLOPS

FP32 163.4 TFLOPS Tensor FP32 163.4 TFLOPS

FP64 81.72 TFLOPS Tensor FP64 163.4 TFLOPS

Tensor INT8 2.62 POPS

Texture Rate 2.55 TTexel/s

Core Config

Shaders 19,456 Shaders

TMUs 1,216 TMUs

Tensor Cores 1,216 T-Cores

CUs 304 CUs

Clocks

Base Clock 1GHz

Boost Clock 2.1GHz

Cache

L1 32KB/CU

L2 Cache 32MB shared

L3 Cache 256MB shared

Memory

256GB HBM3e

Memory Bus 8192-bit

Memory Speed 6GT/s

Memory Bandwidth 6.14TB/s

ECC Yes

Power & Thermals

TDP 1000W

API Support

OpenCL 3

GFX 9.4

Decoder

Decoder 4x VCN 4.0.3

Codec JPEG VP9 AVC (H.264) HEVC (H.265) AV1

Card

Form Factor OAM Module

PCIe PCIe 5.0 x16

Cooling Passive

Multi-GPU Multi-GPU Support Supported Multi-GPU Type Infinity Fabric

Chip

Manufacturer AMD

Chip Designer AMD

Architecture CDNA 3

Family Instinct

Branding

Codename Aqua Vanjaram

Chip Variant Aqua Vanjaram XTX

Market Segment Server

Release Date Oct 10, 2024

Fabrication

Foundry TSMC

Other Foundries TSMCIO

Fabrication Node N5 N6IO

Die Size 920mm² 1480mm²IO

Transistor Count 104B 48.8BIO

Transistor Density 113 MTr/mm² 32.97 MTr/mm²IO

Images