CDNA 3

AMD Instinct MI300X

192 GB HBM3 challenger for large-scale AI training

VRAM

192 GB

Bandwidth

5.3 TB/s

FP16

1,307.4 TFLOPS

TDP

750W

AMD Instinct MI300X

Technical Specifications

VRAM 192 GB HBM3
Memory Bandwidth 5.3 TB/s
FP16 Performance 1,307.4 TFLOPS
BF16 Performance 1,307.4 TFLOPS
FP32 Performance 163.4 TFLOPS
INT8 Performance 2,614.9 TOPS
TDP 750W
Form Factor OAM (Open Accelerator Module)
Interconnect AMD Infinity Fabric (896 GB/s)
Max GPUs per Server 8 (Universal Baseboard)

Prices vary with supply and import costs. Contact for current India pricing.

Best For

LLM inference for the largest models (LLaMA 405B, Mixtral) without model parallelism
Training workloads where VRAM capacity is the bottleneck
Organizations exploring an NVIDIA alternative with ROCm ecosystem
Research labs that need 192 GB per GPU for experimental models

Not Ideal For

  • Teams heavily dependent on CUDA-only libraries
  • Inference deployments where power efficiency matters (750W TDP is high)

Overview

The AMD Instinct MI300X is AMD's flagship data center GPU, featuring 192 GB of HBM3 memory with 5.3 TB/s bandwidth. This is the largest memory capacity of any single GPU available, making it uniquely suited for workloads that are memory-bound.

For LLM inference, the MI300X can serve LLaMA 70B at full FP16 precision on a single GPU, and LLaMA 405B across just two GPUs. Compare this to NVIDIA H100 which requires model parallelism across 4-8 GPUs for the same models. This simplifies deployment and reduces interconnect overhead.

The MI300X runs on AMD's ROCm software stack, which has made significant progress in supporting PyTorch, JAX, and popular inference frameworks. While CUDA ecosystem maturity remains an advantage for NVIDIA, teams that have validated their workloads on ROCm can benefit from competitive pricing and availability.

Get AMD Instinct MI300X pricing for your setup

Tell us your workload and cluster size. We'll quote the complete solution including servers, networking, and colocation.