One year ago, we announced the creation of the IBM Research AI Hardware Center, a global research hub headquartered in Albany, New York. Building on work of the last few years, the launch of the Center initiated the next phase in a long-term effort to combine evolving, fundamental advances in AI with new computing accelerators, technologies, and architectures designed and optimized specifically for AI computation. Launching the Center was our way to invite industrial and academic partners to join us in the development and productization of AI hardware, essential to meeting the computational demands of the emerging broad AI era. Our inherently energy efficient technologies aim to meet the demand for computational power that AI applications and software advances require.
The Path to 1000x
The Center is taking a holistic, end-to-end approach to AI hardware, spanning the computing stack all the way from fundamental materials research to AI workloads. Our goal is to improve compute performance efficiency 1,000x by 2029, with the annual delivery of new AI accelerator cores, targeting 2.5X improvement per year. We’ve more than doubled that gain in the first year.
The Center is working on several projects in tandem that will help us reach this goal. We are a global leader in approximate computing architecture and software techniques for deep learning with our Digital AI Cores. The Center also develops Analog AI Cores, paving a path to ideal analog core behavior by combining materials, device, and algorithmic innovations. Other innovations in heterogeneous integration chip packaging aim to eliminate memory bandwidth bottlenecks. And, we are creating a software ecosystem for AI hardware tools, as well as an AI testbed environment and ecosystem for AI tool exploration and assessing workloads.
We are now joined by 10 enterprise and academic institutions in our AI Hardware Center. As part of our AI testbed, members have access to Rennslaer Polytechnic Institute’s AiMOS supercomputer — one of the most powerful — and energy-efficient — supercomputers in the world as well as technologies created, developed, tested, and matured in our unparalleled research cleanroom facility in Albany, NY.
Our members collaborate with us in a variety of ways. They can use early access to new hardware technologies and our emulation platform to develop application software before hardware is commercialized. Some of our members are testing our AI for manufacturing automated templates to create custom optimized models for manufacturing controls. Other members use IBM AI tools to create new AI models for enterprise workloads, assessing accuracy and performance. With our equipment, consumable and foundry partners, we are perfecting process technologies for new analog device elements and building testchips. Likewise, we are developing and testing 3D packaging technologies and consumables to eliminate memory bandwidth bottlenecks in AI training. We are partnering with IBM Systems and IBM Cloud teams to bring our AI hardware technologies to our IBM clients.
Year One Results
AI Hardware Center researchers in the Albany hub and global network have already leveraged the Center for research breakthroughs. For example, researchers have published papers ultra-low precision training and inference of deep neural networks and analog materials advances. The IBM Journal of Research & Development focused its fourth quarter 2019 issue on AI hardware advances.
Our outlook for 2020 is bright. We anticipate new members joining the Center throughout the year. Many of our latest innovations are coming together in new 2020 Digital and Analog AI testchips. We will show our progress against our ambitious roadmap to improve compute performance efficiency 1000x, with increasingly complex workloads. In 2020 we will see our technologies become central to improving AI technology’s environmental sustainability by supporting expanding AI workloads while reducing carbon footprint. The new AiMOS supercomputer at RPI will fuel rapid progress on AI core technology and AI hardware. If you think computers and AI have dramatically changed the world, wait until you see what’s next.
Today, as reported in Nature Machine Intelligence, my colleagues and I have demonstrated a novel approach to deep learning that incorporates biologically-inspired neural dynamics and enables in-memory acceleration, bringing it closer to the way in which the human brain works. The results point towards the broad adoption of more biologically-realistic deep learning for applications in […]
Scientists around the world are inspired by the brain and strive to mimic its abilities in the development of technology. Our research team at IBM Research Europe in Zurich shares this fascination and took inspiration from the cerebral attributes of neuronal circuits like hyperdimensionality to create a novel in-memory hyperdimensional computing system. The most […]
Can analog AI hardware support deep learning inference without compromising accuracy? Our research team at IBM Research Europe in Zurich thought so when we started developing a groundbreaking technique that achieves both energy efficiency and high accuracy on deep neural network computations using phase-change memory devices. We believe this could be a way forward in […]