A $5 million National Science Foundation (NSF) award will allow the Pittsburgh Supercomputing Center (PSC) to deploy a unique high-performance artificial intelligence (AI) system. Neocortex will introduce fundamentally new hardware to greatly speed AI research. PSC, a joint research organization of Carnegie Mellon University and the University of Pittsburgh, will build the new supercomputer in partnership with Cerebras Systems and Hewlett Packard Enterprise (HPE).
Neocortex will go beyond the technologies that have powered much of the advancement in AI since 2012. The system will do this by exploring a revolutionary combination of Cerebras Wafer Scale Engine (WSE) processors, which are designed specifically to accelerate AI, and an extremely large-memory HPE Superdome Flex system for massive data handling capability. This balanced system will democratize access for researchers to game-changing compute power for training, the most time-consuming step of AI, to be much faster, even interactive. Neocortex is planned to be available at no cost to researchers nationwide starting later in 2020.
The Neocortex project is led by Paola Buitrago, Principal Investigator and PSC Director of Artificial Intelligence and Big Data. The system’s unique AI capabilities for advanced research and development, and its speed, will remove barriers to AI innovation, supporting the Executive Order on Maintaining American Leadership in Artificial Intelligence and the American Artificial Intelligence Initiative.
“Neocortex is a first of its kind in many ways,” Buitrago said. “With Neocortex, we will be closer to achieving our vision of interactive AI development by offering capabilities that are truly revolutionary. It’s our privilege to offer this outstanding resource at no cost to the research community. We look forward to all the great discoveries it will enable.”
“These awards represent a suite of complementary advanced computational capabilities and services aimed to empower new fundamental research in many fields,” said Amy Friedlander, Acting Director of NSF’s Office of Advanced Cyberinfrastructure. “NSF’s long-standing investments in advanced and innovative computing respond to the rapid evolution and expansion of computational- and data-intensive research being conducted across all of science and engineering.”
Deep Learning at the Speed of Thought
A field of AI called deep learning is revolutionizing science, medicine and the way we live. Deep learning allows for faster, more accurate forecasts of severe weather; advances in precision medicine, by tailoring treatments to individual patients; the voice assistants in our phones and homes; and many other applications in science, medicine, technology and industry. This AI approach also helps scientists find rare events in Big Data and speeds simulations to run up to billions of times faster, accelerating the time to scientific results and saving energy. Deep learning now crosscuts research in many fields, and its use is rapidly expanding to others.
Applying deep learning starts with training, or applying iterative optimization algorithms to tune deep neural networks by using collections of data. The networks are deep because they contain many layers of artificial “neurons,” and, optimizing the many connections―typically ranging from tens of millions to billions―between these neurons across and within layers requires massive amounts of computing. Training is that optimization process, and can take a very long time: days, weeks or even months. Long training times limit the scope and complexity of the scientific challenges that can be addressed, and they prevent researchers from exploring promising approaches.
Graphics processing units (GPUs) have dominated AI since 2012, when researchers discovered that GPUs greatly accelerated deep learning compared to the alternatives then available. But what if training could be made even faster, by tens to a thousand times? The result would be transformative: scientists and engineers could rapidly develop and refine their ideas, enabling them to achieve high-impact solutions to the most pressing and complex issues. That is the goal of Neocortex.
AI for Discovery and Societal Good
AI is transforming our ability to find information in complex data and to scale simulations to predict phenomena at large scale. Such advances are likely to help solve many of our most urgent problems, including quickly understanding and suppressing emerging pandemics, realizing precision medicine to improve quality of life and reduce healthcare costs, understanding the biology of rare diseases to open the door to treatments and reducing our impact on the environment through improving the efficiencies of renewable power generation and transportation. Neocortex, and the advanced support that PI Buitrago will provide for the system through PSC’s AI & Big Data group, will accelerate these fields and many more.
The Neocortex project will additionally focus on building a strong community around its revolutionary capabilities, including collaborations with other leading national institutions and emphasizing inclusion and diversity. It will build STEM talent through training and internships, develop the U.S. workforce and national competitiveness through industrial outreach, and foster international collaborations.
Revolutionary Processors and Massive Memory
The novel Neocortex architecture will transform deep learning research by coupling two exceptionally powerful Cerebras CS-1 AI servers with an extreme shared-memory HPE Superdome Flex server to achieve unprecedented AI scalability with excellent system balance.
“We are immensely proud to be a part of the game-changing introduction of Neocortex, which leverages the massive computational power of the CS-1 to advance AI research,” said Andrew Feldman, CEO and co-founder of Cerebras. “We invented CS-1 to be the industry’s most powerful AI computer, and when coupled with HPE’s advanced memory server, it can truly accelerate and improve the future of science research.”
Each Cerebras CS-1 is powered by one Cerebras Wafer Scale Engine (WSE) processor, a revolutionary high-performance processor designed specifically to accelerate deep learning training and inferencing. The Cerebras WSE is the largest computer chip ever built, containing 400,000 AI-optimized cores implemented on a 46,225 square millimeter wafer with 1.2 trillion transistors, compared to only billions of transistors in high end CPUs and GPUs.
Neocortex will use the HPE Superdome Flex, an extremely powerful, user-friendly front-end high-performance computing (HPC) solution for the Cerebras CS-1 servers. This will enable flexible pre- and post-processing of data flowing in and out of the attached WSEs, preventing bottlenecks and taking full advantage of the WSE capability. HPE Superdome Flex will be robustly provisioned with 24 terabytes of memory, 205 terabytes of high-performance flash storage, 32 powerful Intel Xeon CPUs, and 24 network interface cards for 1.2 terabits per second of data bandwidth to each Cerebras CS-1.
“HPE has a long-standing collaboration with PSC to develop powerful joint solutions, which includes coupling the world’s first converged HPC and AI supercomputer in PSC’s Bridges to deliver breakthroughs in science and medicine,” said Mike Woodacre, HPE Fellow, CTO, HPC & MCS, at HPE. “We are honored to be a part of a uniquely new collaboration on Neocortex, combining next-generation HPE Superdome Flex in-memory processing with Cerebras CS-1 to enable researchers to tackle complex AI workloads at unprecedented speed.”
Neocortex will support the most popular deep learning frameworks and automatically and transparently accelerate them, providing researchers with great ease of use. It will be federated with PSC’s new Bridges-2 supercomputer, also to be installed in 2020, to form a singularly powerful and flexible ecosystem for high performance AI, data analytics, modeling and simulation.
If you want to learn more about Neocortex, please email us at firstname.lastname@example.org.