Skip to main content

AMD ROCm / GPU

Vendor: AMD

Category: GPU Graphics Processor

Architecture: CDNA 4 / CDNA 3 / CDNA 2

Introduction​

AMD Radeon graphics and Instinct compute card GPU computing platform. ROCm (Radeon Open Compute) is AMD's open-source GPU computing platform, supporting Radeon RX, Radeon PRO, and Instinct MI series.

Specifications​

ModelComputeMemoryInterfaceTDPProcess
MI350XTBA (FP8)288GB HBM3EOAMTBA3nm (CDNA 4)
MI325X2,614 TFLOPS (FP8) / 1,307 (FP16)256GB HBM3EOAM750W5nm (CDNA 3)
MI300X2,614 TFLOPS (FP8) / 1,307 (FP16)192GB HBM3OAM750W5nm (CDNA 3)
MI300A983 TFLOPS (FP8) / 246 (FP16)128GB HBM3OAM (CPU+GPU)750W5nm (CDNA 3)
MI250X383 TFLOPS (FP16)128GB HBM2eOAM560W6nm (CDNA 2)
RX 7900 XTX61 TFLOPS (FP32)24GB GDDR6PCIe 4.0355WRDNA 3

Official Website​

Visit Official Website

Driver Downloads​

Windows​

Linux​

OS Support​

WindowsLinuxmacOSAndroid
⚠️ (Partial)✅❌❌

Version History​

VersionRelease DateDescription
ROCm 6.12024-Q4MI300X full support
ROCm 6.02024-Q1PyTorch 2.x native integration
ROCm 5.72023-Q3CDNA 3 support

Performance Benchmarks​

ModelTaskPerformance Metric
MI300X × 8GPT-3 175B Training~1.5 days (MLPerf)
MI300XLlama 2 70B Inference~95 tok/s (FP16)
MI250X × 8BERT-Large Training~85% H100 efficiency
RX 7900 XTXStable Diffusion XL~2.2s/img

Pricing​

ModelReference PriceNotes
MI300X$12,000-16,000Better cost-performance than H100
MI250X$8,000-12,000Previous-gen data center card
RX 7900 XTX¥4,099-4,999Consumer flagship

Quick Installation​

Linux (Ubuntu 22.04)​

# 1. Add ROCm repository
wget https://repo.radeon.com/amdgpu-install/6.1/ubuntu/jammy/amdgpu-install_6.1.60100-1_all.deb
sudo dpkg -i amdgpu-install_6.1.60100-1_all.deb
sudo amdgpu-install --usecase=rocm

# 2. Verify installation
rocminfo
/opt/rocm/bin/rocm-smi

ROCm currently officially supports Linux only. Windows support is in preview stage.

Code Examples​

Python (PyTorch ROCm)​

import torch

# ROCm uses HIP backend, API consistent with CUDA
assert torch.cuda.is_available(), "ROCm GPU not found"
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"HIP version: {torch.version.hip}")

# PyTorch ROCm version can directly run CUDA code
x = torch.randn(2048, 2048).cuda()
y = torch.matmul(x, x)
print(f"HIP matrix multiply: {y.shape}")

PyTorch ROCm version must be downloaded separately from pytorch.org, selecting the ROCm version.

Architecture Highlights​

  • CDNA 3 Architecture (MI300): Chiplet design, CPU+GPU packaged on the same substrate (MI300A); Infinity Fabric interconnect
  • Open-Source Strategy: ROCm is fully open-source, including compute runtime, compiler, and libraries
  • HIP Compatibility Layer: HIP API is highly compatible with CUDA, requiring minimal changes to port CUDA code

Model Compatibility​

Model/FrameworkSupportNotes
PyTorch✅ NativeROCm backend, CUDA-API compatible
TensorFlow⚠️ LimitedVia PluggableDevice support
JAX✅Via plugin support
Llama and similar LLMs✅vLLM / llama.cpp both support ROCm
Stable Diffusion✅DirectML / ROCm backend
Whisper✅OpenAI Whisper native support

Large-Scale Cluster Deployments​

Based on global AI supercomputing cluster statistics, AMD ROCm has accumulated over 145,952 chips deployed across 17 publicly disclosed clusters.

Chip Model Statistics​

Chip ModelTotal DeployedCluster Count
AMD Radeon Instinct MI250X68,6727
AMD Instinct MI300A51,6965
AMD Instinct MI300X20,3843
AMD Instinct MI2105,2002

Notable Deployment Clusters Top 10​

#Cluster NameTotal ChipsChip ModelOperator
1Lawrence Livermore NL El Capitan Phase 244,544AMD Instinct MI300A ×44,544US Department of Energy, United States of America
2Oak Ridge NL Frontier37,632AMD Radeon Instinct MI250X ×37,632US Department of Energy, United States of America
3Oracle OCI MI300x16,384AMD Instinct MI300X ×16,384Oracle, United States of America
4Eni HPC613,888AMD Radeon Instinct MI250X ×13,888Eni, Italy
5EuroHPC LUMI11,912AMD Radeon Instinct MI250X ×11,912EuroHPC JU, Finland
6Lawrence Livermore NL Tuolumne4,608AMD Instinct MI300A ×4,608US Department of Energy, United States of America
7Core42 AI-034,320AMD Instinct MI210 ×4,320G42, United Arab Emirates
8Vultr Chicago Cluster3,000AMD Instinct MI300X ×3,000Vultr, United States of America
9Microsoft Explorer-WUS31,920AMD Radeon Instinct MI250X ×1,920Microsoft, United States of America
10Sandia NL El Dorado1,520AMD Instinct MI300A ×1,520US Department of Energy, United States of America

If you're evaluating alternatives, the following products may also fit your scenario: