This article was accepted into the corpus but its outbound wikilinks were never NER-processed — typical at the deepest BFS hop or when the run's entity cap was reached. No expansion funnel to show.
| HIP (Heterogeneous-Compute Interface for Portability) | |
|---|---|
| Name | HIP (Heterogeneous-Compute Interface for Portability) |
| Developer | Advanced Micro Devices |
| Initial release | 2015 |
| Latest release | 2024 |
| Repository | ROCm |
| License | MIT-like |
| Platform | Linux |
HIP (Heterogeneous-Compute Interface for Portability) is a C++ runtime and programming model designed for portable development of parallel workloads on heterogeneous processors. Created to assist migration between vendor-specific platforms, HIP aims to enable source-level portability for compute kernels across multiple accelerator vendors and hardware families. It integrates with compiler toolchains and driver stacks to map parallel kernels to device-specific runtimes.
HIP provides a C++ dialect and runtime that exposes kernel launch syntax, memory management, and synchronization primitives for accelerators. It targets interoperability with existing ecosystems such as AMD, NVIDIA, Intel Corporation, Microsoft, Google, Apple Inc. and integrates with projects including LLVM Project, Clang, GNU Compiler Collection, Linux Foundation, Khronos Group, and OpenMP. The project is distributed via the ROCm software stack and is maintained alongside efforts from organizations like The Linux Foundation and research groups at institutions such as Massachusetts Institute of Technology, Stanford University, and University of California, Berkeley.
HIP originated within Advanced Micro Devices engineering groups during an era of growing interest in heterogeneous computing exemplified by events like the International Supercomputing Conference and initiatives such as the National Science Foundation programs for exascale research. Its public trajectory coincided with AMD promoting the ROCm initiative and engaging with open-source communities including contributors from Canonical (company), Red Hat, SUSE, and academia. Over time HIP incorporated work from the LLVM Project and saw contributions from engineers with backgrounds at NVIDIA Corporation, Intel Corporation, and cloud providers like Amazon Web Services, Microsoft Azure, and Google Cloud Platform.
Key milestones include initial ROCm releases supporting discrete GPUs, upstreaming of toolchain components into LLVM Project and Clang repositories, and collaborations around standards bodies such as the Khronos Group and initiatives involving the OpenACC and OpenMP communities. The design choices reflected lessons from earlier platforms, including concepts from CUDA (computing platform), OpenCL, and vendor SDKs developed by NVIDIA and Intel Corporation.
HIP exposes an API that mirrors many kernel and memory semantics familiar to developers from platforms like CUDA (computing platform), while mapping at compile- or link-time to backends for vendors such as AMD and NVIDIA. The programming model includes concepts of kernels, device memory allocation, streams, events, and peer-to-peer communication, enabling integration with frameworks like TensorFlow, PyTorch, Caffe, Theano, and numerical libraries originating from groups at Lawrence Livermore National Laboratory and Sandia National Laboratories.
Compilers based on LLVM Project transform HIP source into either AMD GPU code within the ROCm stack or CUDA-compatible code paths for NVIDIA Corporation devices. Runtime components interact with kernel dispatch layers similar to those in Mesa (software) and driver architecture maintained by X.Org Foundation contributors. HIP supports interoperability with MPI implementations from projects such as Open MPI and MPICH, and with linear algebra libraries like BLAS, LAPACK, and vendor-tuned libraries from Intel Corporation and Advanced Micro Devices.
The ecosystem around HIP includes compiler frontends derived from Clang, debuggers and profilers comparable to GDB, Valgrind, NVIDIA Nsight, and performance tools developed by teams at Argonne National Laboratory and Oak Ridge National Laboratory. Packaging and distribution are integrated with Linux distributions maintained by Canonical (company), Red Hat, and SUSE, and with continuous integration systems used by projects hosted on GitHub, GitLab, and archives like SourceForge.
HIP supports bindings and language interop layers enabling use from Python (programming language) via projects like NumPy, SciPy, CuPy, and deep learning frameworks such as PyTorch and TensorFlow. Vendor-supplied numerical libraries and community projects provide optimized kernels for workloads typical in research from institutions like Lawrence Berkeley National Laboratory and industry teams at Google, Facebook, and Microsoft Research.
Performance characteristics of HIP depend on the quality of backend mappings and driver support from vendors including Advanced Micro Devices and NVIDIA Corporation. Benchmarks reported by organizations such as Top500 participants, research groups at Oak Ridge National Laboratory, and industrial teams at Intel Corporation demonstrate parity in many dense linear algebra and convolutional workloads when compiler optimizations and vendor libraries are tuned. Compatibility layers enable compilation to either native AMD GPU code via ROCm or translation layers targeting CUDA (computing platform); however, microarchitectural features of devices—documented in papers from IEEE conferences and presentations at International Conference on Supercomputing—impact peak throughput and latency behavior.
HIP has been adopted in high-performance computing centers like Oak Ridge National Laboratory, Lawrence Livermore National Laboratory, and Lawrence Berkeley National Laboratory for porting scientific codes and simulation frameworks developed at Los Alamos National Laboratory. Industry adopters include cloud providers such as Amazon Web Services, Microsoft Azure, and Google Cloud Platform that integrate GPU-accelerated services, and enterprises in domains represented at conferences like NeurIPS, ICML, and CVPR for machine learning workloads. Academic projects from Massachusetts Institute of Technology, Stanford University, and University of California, Berkeley use HIP to target heterogeneous clusters in research on computational physics, computational chemistry, and data analytics.
Compared with CUDA (computing platform), HIP emphasizes source portability and open-source toolchain integration through ROCm and LLVM Project components, whereas CUDA remains closely integrated with NVIDIA Corporation hardware and tooling. Alternatives such as OpenCL and standards promoted by the Khronos Group trade different portability and performance guarantees; industry efforts like SYCL from Khronos Group and vendor initiatives from Intel Corporation present other cross-platform approaches. Choice between HIP, CUDA, OpenCL, and SYCL often depends on target hardware from Advanced Micro Devices or NVIDIA Corporation, development resources at organizations like IBM and Microsoft Research, and project histories tied to ecosystems such as TensorFlow and PyTorch.
Category:Programming languages Category:Parallel computing Category:Advanced Micro Devices