OSC offers GPU computing on all its systems. While GPUs can provide a significant boost in performance for some applications, the computing model is very different from the CPU. This page will discuss some of the ways you can use GPU computing at OSC.
To request nodes with a GPU add the --gpus-per-node=x
attribute to the directive in your batch script, for example, on Owens:
#SBATCH --gpus-per-node=1
In most cases you'll need to load the cuda module (module load cuda
) to make the necessary Nvidia libraries available.
The GPUs on Owens and Pitzer can be set to different compute modes as listed here. They can be set by adding the following to the GPU specification when using the srun
command. By default it is set to shared
.
srun --gpu_cmode=exclusive
or
srun --gpu_cmode=shared
The compute mode shared
is the default on GPU nodes if a compute mode is not specified. With this compute mode, mulitple CUDA processes on the same GPU device are allowed.
We have several supported applications that can use GPUs. This includes
Please see the software pages for each application. They have different levels of support for multi-node jobs, cpu/gpu work sharing, and environment set-up.
There are a few libraries that provide GPU implementations of commonly used routines. While they mostly hide the details of using a GPU there are still some GPU specifics you'll need to be aware of, e.g. device initialization, threading, and memory allocation. These are available at OSC:
MAGMA is an implementation of BLAS and LAPACK with multi-core (SMP) and GPU support. There are some differences in the API of standard BLAS and LAPACK.
cuBLAS is a highly optimized BLAS from NVIDIA. There are a few versions of this library, from very GPU-specific to nearly transparent. cuSPARSE is a BLAS-like library for sparse matrices.
The MAGMA library is built on cuBLAS.
cuFFT is NVIDIA's Fourier transform library with an API similar to FFTW.
cuDNN is NVIDIA's Deep Neural Network machine learning library. Many ML applications are built on cuDNN.
GPUs present a different programming model from CPUs so there is a significant time investment in going this route.
OpenACC is a directives-based model similar to OpenMP. Currently this is only supported by the Portland Group C/C++ and Fortran compilers.
OpenCL is a set of libraries and C/C++ compiler extensions supporting GPUs (NVIDIA and AMD) and other hardware accelerators. The CUDA module provides an OpenCL library.
Our GPUs span several generations with different capabilites and ease-of-use. Many of the differences won't be visible when using applications or libraries, but some features and applications may not be supported on the older models.
The P100 "Pascal" is a NVIDIA GPU with a compute capability of 6.0. The 6.0 capability includes unified shared CPU/GPU memory -- the GPU now has its own virtual memory capability and can map CPU memory into its address space.
Each P100 has 16GB of on-board memory and there is one GPU per GPU node.
The NVIDIA V100 "Volta" GPU, with a compute capability of 7.0, offers several advanced features, one of which is its Tensor Cores. These Tensor Cores empower the GPU to perform mixed-precision matrix operations, significantly enhancing its efficiency for deep learning workloads and expediting tasks such as AI model training and inference.
The V100 deployed in 2018 comes equipped with 16GB of memory, whereas the V100 deployed in 2020 features 32GB of memory. There are two GPUs per GPU node,
Additionally, there are four large memory nodes equipped with quad NVIDIA Volta V100s with 32GB of GPU memory and NVLink.
The NVIDIA A100 "Ampere" GPU, with a compute capability of 8.0, empowers advanced deep learning and scientific computing tasks. For instance, it accelerates and enhances the training of deep neural networks, enabling the training of intricate models like GPT-4 in significantly less time when compared to earlier GPU architectures.
The A100 comes equipped with 80GB of memory. here are 4 GPUs with NVLink, offering 320GB of usable GPU memory per node.