ROCm
AMD's open-source GPU compute platform — the AMD equivalent of CUDA for running AI workloads on Radeon GPUs.
ROCm (Radeon Open Compute) is AMD's open-source GPU compute stack. It's the software platform that allows AI frameworks like PyTorch to run on AMD Radeon GPUs — AMD's answer to CUDA. For local AI users considering AMD hardware, ROCm compatibility is the most important factor to evaluate before purchasing.
What ROCm Enables
Without ROCm support, AMD GPUs are limited to CPU offload or Metal (on macOS). With ROCm, you can run full GPU-accelerated inference through llama.cpp (ROCm build), PyTorch, and Ollama on Linux. ROCm-accelerated inference on an RX 7900 XTX (24GB VRAM) is dramatically faster than CPU-only and competitive with NVIDIA at the same VRAM tier.
The key limitation: ROCm is primarily a Linux story. Windows support through HIP (ROCm's CUDA compatibility layer) is improving but remains less stable. If you're running local AI on Windows, NVIDIA is still the path of least resistance.
Compatible Hardware
Not all AMD GPUs support ROCm. The officially supported list for ROCm 6.x includes:
- RX 6800, 6900 series (RDNA 2)
- RX 7800, 7900 series (RDNA 3)
- RX 9070, 9070 XT (RDNA 4)
Older RDNA 1 and Polaris cards have community-maintained support but no official guarantees.
ROCm vs. CUDA in Practice
ROCm's main advantages: open-source (AMD publishes the full stack), no licensing restrictions, and AMD's VRAM-per-dollar is often better than NVIDIA at equivalent price points (e.g., RX 7900 XTX offers 24GB for less than an RTX 4090).
The disadvantages: smaller ecosystem, fewer pre-compiled binaries, and some optimizations like specific quantized kernels arrive on ROCm later than CUDA. Software that "just works" on CUDA may require manual compilation on ROCm.
For budget-conscious local AI users comfortable with Linux, AMD + ROCm is a viable and increasingly capable alternative. For maximum compatibility with minimal configuration, NVIDIA remains the default recommendation.
Related guides: AMD vs NVIDIA for local LLMs — full comparison of ROCm versus CUDA across inference speed, software compatibility, and cost per GB of VRAM. AMD Ryzen AI Max VRAM and GTT memory on Linux — how ROCm interacts with AMD's unified memory architecture on Strix Halo hardware.