rocm

ROCm (Radeon Open Compute) is AMD's open-source GPU computing platform for HPC and AI workloads, competing with NVIDIA's CUDA ecosystem. Components: (1) HIP (Heterogeneous-compute Interface for Portability)—CUDA-like C++ API enabling code to run on both AMD and NVIDIA GPUs, (2) rocBLAS/rocFFT/rocSOLVER (math libraries equivalent to cuBLAS/cuFFT), (3) MIOpen (deep learning primitives like cuDNN), (4) RCCL (collective communications like NCCL). Hardware: AMD Instinct MI250X/MI300X accelerators targeting data center AI training. PyTorch support: native ROCm backend (torch.cuda works via HIP translation). Advantages: open-source stack, competitive hardware performance (MI300X: 192GB HBM3), and cost alternative to NVIDIA. Challenges: smaller ecosystem, fewer optimized libraries, less community tooling, and compatibility gaps with CUDA-specific features. hipify tool: semi-automated CUDA-to-HIP code conversion. Growing adoption: major cloud providers (Azure, AWS) offering AMD GPU instances. Key for breaking NVIDIA monopoly in AI training infrastructure.