← AI Terminology

ROCm - Radeon Open Compute

ROCm (Radeon Open Compute) is AMD's open-source GPU compute platform — a software stack providing compilers, runtime, libraries (MIOpen, rocBLAS, hipBLAS), and framework backends (PyTorch-ROCm, TensorFlow-ROCm) that enable deep learning workloads on AMD GPUs.

It is AMD's answer to NVIDIA's CUDA ecosystem.
Why It Matters in AI
NVIDIA's CUDA dominance in AI rests as much on software as hardware — years of optimised libraries, framework integrations, and developer tooling. ROCm is AMD's attempt to break this moat: providing a CUDA-compatible programming model (HIP) and equivalent library stack that lets PyTorch and TensorFlow run on AMD Instinct GPUs. As AMD's MI300X gains traction for LLM inference (it has 192GB HBM3 vs H100's 80GB), ROCm's maturity determines whether AMD can compete for AI workloads beyond price.
Key Points
Aspect Description
HIP Heterogeneous-compute Interface for Portability — CUDA-compatible C++ dialect; hipcc compiler
hipify Tool to automatically convert CUDA code to HIP — port CUDA kernels to AMD GPUs
Libraries MIOpen (cuDNN equiv), rocBLAS (cuBLAS equiv), rocFFT, hipRAND — GPU math primitives
Maturity gap ROCm lags CUDA in kernel tuning, debugging tools, and some library coverage — improving rapidly
PyTorch-ROCm Official PyTorch build with ROCm backend: pip install torch --index-url https://download.pytorch.org/whl/rocm6.0
Hardware support MI300X (192GB HBM3), MI250X, MI210, RX 7900 XTX (consumer) — data centre + consumer GPU AI
Simple Analogy
An alternative operating system for AI hardware: just as Linux offers a CUDA alternative to Windows (NVIDIA's comfort zone), ROCm offers an open-source, HIP-based alternative to CUDA — letting AMD GPUs run the same PyTorch code with minimal changes.
Common Usage Examples
  • pip install torch torchvision --index-url https://download.pytorch.org/whl/rocm6.0 — PyTorch + ROCm
  • import torch; torch.cuda.is_available() — returns True on ROCm (HIP exposes CUDA-compatible interface)
  • hipcc kernel.hip -o kernel — compile HIP kernel for AMD GPU
  • rocm-smi — AMD equivalent of nvidia-smi — monitor GPU utilisation and temperature
  • docker pull rocm/pytorch:latest — AMD's official PyTorch + ROCm Docker image
Summary
In short: ROCm is AMD's open-source GPU compute platform — providing a CUDA-compatible software stack (HIP, MIOpen, PyTorch-ROCm) that enables deep learning workloads on AMD GPUs as they challenge NVIDIA's dominance in AI hardware.