Translate GPU concepts between AMD and NVIDIA terminology. Use when: user asks about AMD using NVIDIA terms (or vice versa), mapping CU to SM, LDS to Shared Memory, XGMI to NVLink, MFMA to Tensor Core, wavefront to warp, HIP to CUDA, ROCm to CUDA Toolkit, MIOpen to cuDNN, RCCL to NCCL, Infinity Cache to L2, OAM to SXM.
Translates GPU architecture and programming concepts between AMD (ROCm/HIP) and NVIDIA (CUDA) terminology.
Load terminology/amd-nvidia.md — it contains complete mapping tables across:
| Section | Covers |
|---|---|
| Architecture Concepts | CU/SM, Stream Processor/CUDA Core, MFMA/Tensor Core, Wavefront/Warp, XCD/Compute Die |
| Memory & Storage | LDS/Shared Memory, VGPR/Register File, Infinity Cache/L2, HBM/HBM, Global Memory |
| Interconnect & Multi-GPU | XGMI/NVLink, Infinity Switch/NVSwitch, OAM/SXM, ROCm RDMA/GPUDirect RDMA, RCCL/NCCL |
| Programming Model | HIP/CUDA, hipcc/nvcc, HIP Runtime API/CUDA Runtime API, Wavefront/Warp |
| Software Ecosystem |
| ROCm/CUDA Toolkit, rocBLAS/cuBLAS, MIOpen/cuDNN, MIGraphX/TensorRT, ROCProfiler/Nsight |
| Key Structural Differences | Wavefront width (64 vs 32), FP64 throughput parity, Infinity Cache vs L2 naming, chiplet design |
Find the relevant table section using the section guide above.
Return the mapping in this format:
Note important asymmetries when they exist:
If the queried concept has no direct counterpart, say so explicitly and explain the closest analog.
For single-term lookups: one-line answer with AMD term ↔ NVIDIA term, followed by a brief description. For multi-term or conceptual queries: a markdown table with columns Generic | AMD | NVIDIA | Notes.