GPU Computing Clusters
|
Learn about the recently released
NVIDIA® C2050 GPU
Computing processors and Tesla™ S2050 1U GPU Computing System here. |
|
GPU (Graphics Processing Unit) Computing Clusters are x86/x86_64 systems that are configured with Graphics Processing Units. This type of cluster utilizes the massive floating point computational power of a modern GPU to perform general purpose computing instead of simply rendering graphics.
At the dawn of GPU computing, GPU cards were specially designed for computer graphics, which made them difficult to program. They lacked standard API capabilities, forcing programmers to adapt video processing methodologies and tools in an effort to meet their computing requirements. This type of implementation is now historically referred to as GPGPU computing.
Today’s high performance computational GPUs are far different than that, and are general-purpose parallel processors with support for accessible programming interfaces and industry-standard languages such as C and Fortran. These more modern implementations are now more often referred to as GPU Computing, although the terms GPGPU and GPU Computing are used interchangeably.
It is also becoming more common today to pair GPUs with general purpose
CPUs in a heterogeneous computing model, with
the sequential part of the application controlled by the CPU, and the
computationally intensive portion on the GPU.
Developers who port their applications to GPUs often achieve
performance increases of orders of magnitude versus optimized CPU only implementations, garnering
lower equipment costs, significantly lower power consumption, and greatly reduced heat load for the same workload. GPU computing is
green
!
So, why isn't everyone already using GPU computing, given the obvious advantages in computing power, lowered power consumption, and decreased
heat load enjoyed by this architecture? It all depends on the software and models you intend to deploy. Are they ported to a GPU model and mature? Does your code run well
in a hybrid GPU computing environment? At this time, as with any early adoption stage of any new technology, circumstances are fluid, and the answers can
change daily. GPU clusters
utilize codes that have been written in or
ported to languages specifically designed to take advantage of GPUs, such as
CUDA
, or
OpenCL
. CUDA is NVIDIA's parallel computing architecture for GPUs, and OpenCL is the new cross-vendor standard for
heterogeneous computing on CUDA and other GPU architectures. NVIDIA has also worked with
Portland Group
to develop a CUDA Fortran Compiler, which is needed to allow many scientific codes to take advantage of GPU computing. NVIDIA also is heavily supporting the OpenCL
standards development process, which will help standardize vendors GPU computing architectures.
Please keep several things in mind as you investigate GPU computing and how it can be used to solve your high performance computing problem, and contact your Aspen Systems sales engineer with your questions as well. They can help you design and deploy a future proofed GPU computing solution that will meet your needs now and in the future.
- GPGPU workloads are best suited for applications that can be improved by a high degree of parallelism.
- While GPU computing implementations are among the most performance and power efficient choices, they are extremely sensitive to their specific software environment, and require a very large degree of optimization to run at their full potential. Early GPU clusters were deployed to run a single or only a few codes that have been specifically designed for the GPU environment, however hybrid clusters who support both general purpose CPU as well as GPU computing are now becoming more prevalent.
-
GPU computing is a rapidly changing eco-system, with
CUDA
and
OpenCL
specifications and capabilities changing rapidly, and new codes ported to GPU computing every day. If you are porting a code to this technology, get involved in the
ever growing GPU computing community at places like
GPGPU.org .
If you intend to utilize a GPU code, check your codes porting status
and maturity on GPUs often, and keep up to date with the changes in your environment as it matures. And, as always,
benchmark your code
on an Aspen Systems benchmark cluster to determine definitive real-time answers, and learn about and utilize the latest
GPU computing technology.
NVIDIA®
Corporation has now released their C2050
Tesla™
GPU Computing Processor. The new C2050, and forthcoming C2070
( available in Q3 ) GPU Computing processors are
based
on the next generation CUDU architecture codenamed
Fermi, and are now an even
better fit for technical and
high performance computing. Read more about the new
NVIDIA® C2050 Tesla™ GPU platform
and
how it can help you solve your high performance computing problems
today. This architecture can often deliver equal computing power as a
purely CPU based
solution at 1/10 the cost and 1/20th the power consumption!
Contact Aspen Systems sales at 1-800-992-9242 for more information about high performance GPU computing systems built with NVIDIA® C2050 GPU Computing processors today!
