AMD MI300X

Data center GPU · CDNA3

Summary

The most widely-available AMD GPU, the MI300X has better specs than the NVIDIA H200, but is generally available at lower prices.

Its additional VRAM and higher memory bandwidth is especially useful for LLM inference applications. Software support for AMD GPUs has historically been less good than for equivalent NVIDIA GPUs, though this has been improving.

For a similar but slightly more powerful GPU, see the AMD MI325X.

Launched
Q2 2024
VRAM
192 GB
Mem. bandwidth
5,200 GB/s
On-demand from
Compare prices from 7 providers →

Tech specs

VRAM 192 GB HBM3
Memory bandwidth 5,200 GB/s
Interface PCIe
TDP 750 W
Supported data types
FP64FP32TF32FP16BF16FP8INT8

Cloud rental prices

7 providers · available in 2 countries · lowest price per GPU, per hour

Prices updated

On-demand
$1.85 – $6.00 /GPU/h
Reserved
$2.29 – $3.85 /GPU/h
Provider On-demand Reserved Spot
Cirrascale $3.08
Crusoe $3.45
DigitalOcean $1.99
Hot Aisle $1.99 $3.39
Oracle $6.00
Seeweb $2.69 $2.29
Vultr $1.85
Compare all AMD MI300X cloud providers & configurations →

Additional providers who may offer AMD MI300X without public pricing: TensorWave, Runpod

Models that fit in VRAM

Open-weights models that fit in VRAM. Estimated using 🤗 accelerate, plus approximation for up to 8K context.

Media

Peak theoretical performance

FP8 Matrix 2,614.9 TFLOPS
INT8 Matrix 2,614.9 TOPS
BF16 Matrix 1,307.4 TFLOPS
FP16 Matrix 1,307.4 TFLOPS
TF32 Matrix 653.7 TFLOPS
FP32 Vector 163.4 TFLOPS
FP32 Matrix 163.4 TFLOPS
FP64 Vector 81.7 TFLOPS
FP64 Matrix 163.4 TFLOPS

Performance figures assume no sparsity; in cases where only sparse performance figures are published by the manufacturer, these are halved to give approximate dense performance.

Resources

Detailed documentation from the manufacturer.

Similar GPUs

Other accelerators you might compare.