1. What's New
Welcome to the 21.7 version of the NVIDIA HPC SDK, a comprehensive suite of compilers and libraries enabling developers to program the entire HPC platform, from the GPU foundation to the CPU and out through the interconnect.
Key features that are new in this release of the NVIDIA HPC SDK for Linux include:
- The 21.7 version of the HPC SDK includes full support for the NVIDIA Arm HPC Developer Kit.
- Performance for Arm CPUs is improved by various enhancements to the HPC Compilers that provide additional intrinsics, and optimized math functions and vectorization.
- Reductions for CUDA Fortran "cuf kernels" can be explicitly specified by the user. Refer to the CUDA Fortran Programming Guide for more details.
- The command-line option -Mint128 turns on support for extended integer types __int128 and unsigned __int128 in NVC and NVC++. Note that 128-bit integer types are not supported in GPU code or in OpenMP or OpenACC. In a future release, 128-bit integer types will be enabled by default.
- The -Minfo option for NVC++ has a new suboption, stdpar. When compiling with -stdpar -Minfo=stdpar the compiler will report whether or not calls to C++ standard algorithms with an execution policy were parallelized. Additionally, when compiling with -stdpar, the compiler will now warn about calls to standard algorithms with a parallel execution policy that are not parallelized.
- The NVIDIA HPC SDK now includes versions 11.4, 11.0, and 10.2 of the CUDA toolchain.
- The behavior of the ieee_arithmetic module's ieee_next_after(x,y) runtime library routine has been updated to be consistent with the Fortran specification and other implementations.
- A new option -gpu=[no]implicitsections has been added to direct the HPC Compilers to [not] implicitly treat array element references in a data clause as an array section. The default behavior for this release matches the behavior of previous releases; the default behavior will change in a future release. Please see the HPC Compilers User Guide or the manpages for more information.
-
For all targets, -O3 now includes two floating
point optimizations that can result in some loss of precision.
These include:
- potentially rewriting floating point division as a multiply by reciprocal (x/y => x*1/y). This behavior can be enabled or disabled by -M[no-]recip-div.
- factorization of floating point types for increased symbolic cancelation. This behavior can be enabled or disabled by -M[no-]factorize.
- On Arm, -O3 and -Mfprelaxed are more aggressive at finding FMA opportunities.
- On Arm and x86, -Mfprelaxed allows more floating point optimization.
- For Fortran codes, -Minline has been improved to increase inlining opportunities.
2. Release Component Versions
The NVIDIA HPC SDK 21.7 release contains the following versions of each component:
| Linux_x86_64 | Linux_ppc64le | Linux_aarch64 | |||||||
|---|---|---|---|---|---|---|---|---|---|
| CUDA 10.2 | CUDA 11.0 | CUDA 11.4 | CUDA 10.2 | CUDA 11.0 | CUDA 11.4 | CUDA 10.2 | CUDA 11.0 | CUDA 11.4 | |
| nvc++ | 21.7 | 21.7 | 21.7 | ||||||
| nvc | 21.7 | 21.7 | 21.7 | ||||||
| nvfortran | 21.7 | 21.7 | 21.7 | ||||||
| nvcc | 10.2.89 | 11.0.228 | 11.4.43 | 10.2.89 | 11.0.228 | 11.4.43 | N/A | 11.0.228 | 11.4.43 |
| NCCL | 2.10.3 | 2.10.3 | 2.10.3 | 2.10.3 | 2.10.3 | 2.10.3 | N/A | 2.10.3 | 2.10.3 |
| NVSHMEM | 2.1.2 | 2.1.2 | 2.1.2 | 2.1.2 | 2.1.2 | 2.1.2 | N/A | N/A | N/A |
| cuBLAS | 10.2.2.89 | 11.2.0.252 | 11.5.4.3 | 10.2.2.89 | 11.2.0.252 | 11.5.4.3 | N/A | 11.2.0.252 | 11.5.4.3 |
| cuFFT | 10.1.2.89 | 10.2.1.245 | 10.5.0.43 | 10.1.2.89 | 10.2.1.245 | 10.5.0.43 | N/A | 10.2.1.245 | 10.5.0.43 |
| cuRAND | 10.1.2.89 | 10.2.1.245 | 10.2.5.43 | 10.1.2.89 | 10.2.1.245 | 10.2.5.43 | N/A | 10.2.1.245 | 10.2.5.43 |
| cuSOLVER | 10.3.0.89 | 10.6.0.245 | 11.2.0.43 | 10.3.0.89 | 10.6.0.245 | 11.2.0.43 | N/A | 10.6.0.245 | 11.2.0.43 |
| cuSPARSE | 10.3.1.89 | 11.1.1.245 | 11.6.0.43 | 10.3.1.89 | 11.1.1.245 | 11.6.0.43 | N/A | 11.1.1.245 | 11.6.0.43 |
| cuTENSOR | 1.3.1 | 1.3.1 | 1.3.1 | 1.3.1 | 1.3.1 | 1.3.1 | N/A | 1.3.1 | 1.3.1 |
| Nsight Compute | 2021.1.1 | 2021.1.1 | 2021.1.1 | ||||||
| Nsight Systems | 2021.2.1.58 | 2021.2.1.58 | 2021.2.1.58 | ||||||
| OpenMPI | 3.1.5 | 3.1.5 | 3.1.5 | ||||||
| HPC-X | N/A | 2.8.1 | 2.8.1 | N/A | 2.8.1 | 2.8.1 | N/A | 2.8.1 | 2.8.1 |
| UCX | N/A | 1.10.0rc1 | 1.10.0rc1 | N/A | 1.10.0rc1 | 1.10.0rc1 | N/A | 1.10.0rc1 | 1.10.0rc1 |
| OpenBLAS | 0.3.13 | 0.3.13 | 0.3.13 | ||||||
| Scalapack | 2.1.0 | 2.1.0 | 2.1.0 | ||||||
| Thrust | 1.9.7 | 1.9.9 | 1.12.0 | 1.9.7 | 1.9.9 | 1.12.0 | 1.9.7 | 1.9.10 | 1.12.0 |
| CUB | N/A | 1.9.9 | 1.12.0 | N/A | 1.9.9 | 1.12.0 | N/A | 1.9.9 | 1.12.0 |
| libcu++ | 1.0.0 | 2.0.0 | 2.0.0 | 1.0.0 | 2.0.0 | 2.0.0 | 1.0.0 | 2.0.0 | 2.0.0 |
3. Supported Platforms
3.1. Platform Requirements for the HPC SDK
| Architecture | Linux Distributions | Minimum gcc/glibc Toolchain | Minimum CUDA Driver |
|---|---|---|---|
| x86_64 |
CentOS 7.2, 7.3, 7.4, 7.5, 7.6, 7.7, 7.8 |
C99: 4.8 |
440.33 |
| ppc64le |
RHEL 7.3, 7.4, 7.5, 7.6, 7.7, 8.0, 8.1 |
C99: 4.8 |
440.33 |
| aarch64 |
RHEL 8.1 |
C99: 4.8 |
450.36 |
3.2. Supported CUDA Toolchain Versions
The NVIDIA HPC SDK uses elements of the CUDA toolchain when building programs for execution with NVIDIA GPUs. Every HPC SDK installation package puts the required CUDA components into an installation directory called [install-prefix]/[arch]/[nvhpc-version]/cuda.
An NVIDIA CUDA GPU device driver must be installed on a system with a GPU before you can run a program compiled for the GPU on that system. The NVIDIA HPC SDK does not contain CUDA Drivers. You must download and install the appropriate CUDA Driver from NVIDIA , including the CUDA Compatibility Platform if that is required.
The nvaccelinfo tool prints the CUDA Driver version in its output. You can use it to find out which version of the CUDA Driver is installed on your system.
- CUDA 10.2
- CUDA 11.0
- CUDA 11.4
4. Known Limitations
- The cuda-gdb debugger is included in this version. Currently, Fortran arrays with non-constant bounds are not handled correctly and querying values will yield incorrect results. Stepping through CUDA Fortran and OpenACC kernels is partially supported, but incorrect line numbers are displayed. For additional general limitations with cuda-gdb, please refer to its documentation.
- When using -stdpar to accelerate C++ parallel algorithms, the algorithm calls cannot include virtual function calls or function calls through a function pointer, cannot use C++ exceptions, can only dereference pointers that point to the heap, and must use random access iterators (raw pointers as iterators work best).
- When nvc++ -stdpar=multicore is used to generate parallel code, OpenMP pragmas in the same translation unit will also be enabled.
5. Deprecations and Changes
- The current default of -gpu=implicitsections will change in a future release to -gpu=noimplicitsections to adhere to the OpenACC specification.
- Starting with the 21.5 version of the NVIDIA HPC SDK, the -cuda option for NVC++ and NVFORTRAN no longer automatically links the NVIDIA GPU math libraries. Please refer to the -cudalib option.
- Support for the Kepler architecture of NVIDIA GPUs was deprecated starting with the 21.3 version of the NVIDIA HPC SDK.
- Support for the KNL architecture of multicore CPUs in the NVIDIA HPC SDK was deprecated in the HPC SDK version 21.3.
Notices
Notice
ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, "MATERIALS") ARE BEING PROVIDED "AS IS." NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE.
Information furnished is believed to be accurate and reliable. However, NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use. No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation. Specifications mentioned in this publication are subject to change without notice. This publication supersedes and replaces all other information previously supplied. NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation.
Trademarks
NVIDIA, the NVIDIA logo, CUDA, CUDA-X, GPUDirect, HPC SDK, NGC, NVIDIA Volta, NVIDIA DGX, NVIDIA Nsight, NVLink, NVSwitch, and Tesla are trademarks and/or registered trademarks of NVIDIA Corporation in the U.S. and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.
