ROCm is AMD's open software stack for GPU-accelerated computing. AMD's own documentation says it "provides the tools needed to program AMD GPUs", and it plays the role for AMD hardware that CUDA plays for NVIDIA.
What is inside ROCm
Per AMD's ROCm documentation, the Core SDK holds the foundational pieces:
- Runtimes and compilers, including HIP, HIPIFY and LLVM.
- Math and compute libraries for deep learning kernels, dense and sparse BLAS, FFT and random number generation.
- Profilers, debuggers, and control and monitoring tools.
The wider ecosystem adds ROCm-enabled HPC applications and deep learning frameworks such as PyTorch.
How it relates to CUDA
Code written for CUDA does not run on ROCm unchanged. HIP is the programming layer that bridges the two, covered in what HIP is in ROCm. For a full side-by-side, read ROCm vs CUDA.
ROCm is one of several answers to the question of what else can run AI work. The rest are mapped in NVIDIA's main competitors in AI.
