Astounding performance increases can be seen with GPU architectures plus a little hard work with helpful developer tools. Moving from somewhat parallel code to a highly parallel architecture requires rethinking of common coding strategies. Problems are divided differently, and coding to the architecture becomes even more crucial than before. Microway customers are at the forefront of GPU success and best practices.
The Northeastern University Research Group is a leader in GPU computing. The GPU team includes 10 PhD-level and 1 post-doc experts focused on a range of GPU research topics. Their scope of research targets applications, tools, and GPU simulation. They are acknowledged leaders in heterogeneous computing platforms and GPU computing on all scale of hardware.
The team will run an Advances in GPU Computing event on October 27, 2011. Hosted by the Northeastern GPU Computing team, this event will share best practices, their latest projects, and more, to attendees. Information can be found at: http://www.ece.neu.edu/GPGPU/Advances/
Professor David Kaeli leads the group and explains a number of their specialties:
A major focus of the group has been the development of a Biomedical Image Analysis Toolbox. This project is a multi-year project supported by the National Science Foundation. The goal is to deliver a set of engineered software GPU libraries in both OpenCL and CUDA to allow the biomedical research community to leverage the power of GPU devices. To date, we have worked on breast cancer imaging, radiotherapy treatment planning, cardio- vascular plaque detection, and 3-D ultrasound imaging. The project includes researchers at Northeastern, BU, RPI, the University of Puerto Rico at Mayaquez, and Massachusetts General Hospital.
A second focus of the team has been the development of software toolsets to allow researchers to easily utilize the massive parallelism provided by GPUs, particularly OpenCL and AMD GPUs. Our initial tool delivers a rich event-based profiling capability – this work was presented at the AMD Fusion Summit this summer. Our second tool is the Caracal project, which allows parallel programs written in CUDA to run on AMD hardware using the Caracal cross-runtime translation environment. This new capability is being highlighted as part of a tutorial being held at the ACM Parallel Architectures and Complication Techniques Conference.
A third focus is on allowing the architecture community to begin to explore new architectural designs using a complete simulation framework. We have developed the Multi2Sim simulation framework, which provides both functional-level and cycle-accurate simulation of the AMD Evergreen architecture. The simulator accepts unmodified OpenCL programs allowing researchers to study the GPU design space. Multi2Sim will also be featured at the ACM Parallel Architecture and Compilation Techniques Conference in a tutorial.
Part of the promise of GPUs is in heterogeneous computing. Run the pieces of your code where appropriate on multiple architectures: large data sets in big system memory, parallel code on the GPUs, and clock speed dependent work on the CPU. Code portability to most devices CPU, GPU, or embedded is the ultimate goal. The OpenCL standard is the beginning:
Our group is working with AMD, NVIDIA and Imagination GPUs. Due to our
leadership in OpenCL computing, we are able to move applications gracefully between different architectures. We are presently developing extensions to OpenCL to allow for significant speedup when programs run on systems containing multiple GPU devices from different vendors.
We are presently working closely with a leading ultrasound imaging manufacturer on an OpenCL-based software architecture for their next generation platform. OpenCL provides the right level of abstraction for a real-time application such as ultrasound, allowing us to leverage CPUs and GPUs to meet the real-time deadlines. Another exciting OpenCL-based application that we recently released as an open-source project is clSurf, which allows the user to perform feature detection and matching independent of pose or scale.
Importantly, the Northeastern group shows demonstrated success through their larger role in the community and applied excellence with the wider academic, medical, research and commercial spheres:
Most of our work is benefitting a much larger community than our own group. We are presently exploring how to utilize multiple GPUs, as well as how to virtualize the massive resources of a single GPU, to provide new levels of parallelism. Our team has also developed a number of memory tuning algorithms that have increased performance by an order of magnitude for
selected applications.
We are designated jointly with Massachusetts General Hospital as an NVIDIA Research Laboratory. Even more important, was our designation as one of three AMD Strategic GPU Academic Research Partners. Our collaboration with AMD recently produced a new text titled: “Heterogeneous Computing with OpenCL”
Our collaborations engaged the biomedical community early on in the New England region, and they are also having an impact on companies working in Homeland Defense technologies. We are now moving into the embedded space, forging additional collaborations with industry leaders such as Analog Devices and Samsung.
Microway hardware and integration expertise have played a role in the development of these successes.
We have worked with Microway on developing demonstrations based on our
clSurf tools to highlight some of the major benefits of Microway’s GPU-based offerings. We are also a customer of Microway, and recognize their ability to provide very good cost/performance in the GPU systems space.
Microway is a co-sponsor of the Advances in GPU Computing event at Northeastern on October 27.
Michael Fried, Microway’s GPGPU Business Manager, will deliver a lecture on “Developing portable, high performance code with Microway’s OpenCL tools and services.” This lecture will address the challenges associated with programming in the GPGPU world and how Microway tools help users develop and test GPGPU code to meet those challenges across different CPUs GPUs.
More information on the Advances in GPU Computing event can be found at: http://www.ece.neu.edu/GPGPU/Advances/
The event will include talks from industry leaders including, AMD, Microway, HP and Analog Devices. Topics will include the spectrum from embedded platforms to supercomputing.
Microway experts can be reached at 508-746-7341 or via email at [email protected] . View Microway HPC products at http://www.microway.com, SC11 Booth 2606, or email [email protected] for a quote.