Garth Gibson is the president and chief executive officer of the Vector Institute, funded by the government of Canada through the CIFAR Pan-Canadian AI Strategy, the province of Ontario, and Canadian industry sponsors.
Ron Bodkin is vice-president of AI engineering and chief information officer at the Vector Institute and engineering lead at the Schwartz Reisman Institute for Technology and Society.
Artificial intelligence has a bottleneck. Demand for compute – shorthand for computing power – is rising, and countries with AI strategies are only now awakening to the importance of addressing it. To help them do so, the Organization for Economic Co-operation and Development recently created a task force charged with providing benchmarks for AI compute capacity to national governments, and has invited Vector to join and contribute. We intend to describe how Canada is set up to solve the problem.
The federal government dedicated $40-million in its new budget to the procurement of infrastructure for researchers working with Canada’s three national AI institutes. This investment, combined with elements of the CIFAR Pan-Canadian AI Strategy, sets the stage to address AI compute in a way that meets the needs of AI researchers and moves forward on national priorities. It’s an approach we’re primed to execute, and one that other countries can emulate.
The cause and consequence of soaring AI compute demand
First, understand the demand for compute. Behind AI’s rising ubiquity and stunning outcomes are models that are quickly getting bigger and more complex. Consider progress in natural language processing, a branch of AI focused on training models to use and understand language. Last June, OpenAI unveiled GPT-3, touted as the largest language model to date with 175 billion parameters, the individual values that compute a model’s output. Observers estimated that training this model would cost more than $3.5-million, even when using an economical cloud computing option. This expense is a stretch for research institutions and companies without huge research and development budgets. Less than a year after GPT-3′s release, Google released a language model 10 times more complex, with 1.6 trillion parameters.
Artificial intelligence (AI) is a term dating back to the 1950s, when it was defined as “the science and engineering of making intelligent machines.” When we talk about AI, we’re usually talking broadly about the concept of machines that can develop abilities usually associated with human intelligence, such as learning a language or spontaneously navigating a road when driving a car.
Machine learning refers to various methods for developing AI. Many computers function by being explicitly programmed to follow a specific series of instructions; machine learning, in contrast, attempts to train a computer to complete tasks or solve problems on its own. People have been working on this for decades. But only in recent years have computers been fast enough – and able to process large enough sets of information – to make more complex developments in machine learning possible.
From our vantage point leading a top AI institute, we can confidently say: This trend isn’t slowing down soon.
Its effect may be to concentrate research on state-of-the-art AI models in a handful of big, global companies. While these companies create real value, we don’t want a scenario in which the best minds in AI, looking for opportunities to work at the leading edge, find that large companies are the only game in town. Progress on national priorities depends on researchers and academic institutions, large and small, having continued access to the AI compute they need.
Public-sector investment in AI compute provides that access. But how AI compute is acquired and then distributed matters, and here in Canada, we’ve set up an approach that can blend the addition of AI compute with a national AI strategy so that new capacity translates into progress. Here are the elements of that approach:
Leverage domestic AI expertise when expanding infrastructure
Countries need to be thoughtful about AI compute procurement. There’s no blueprint, but there’s also no shortage of suggestions. Computing providers and advanced computing institutes around the world will likely be eager to share opinions on how to proceed. It’s important to resist the temptation to simply buy into such suggestions wholesale. Instead, we must ensure domestic AI researchers are deeply involved in the choice and development of the infrastructure so that it properly addresses their needs. It’s also important to not simply default to a commercial cloud solution. Cloud computing is often progressive and future-looking, but it’s not necessarily the most pragmatic or cost-effective research option, especially for rapidly changing technologies such as AI. For countries with AI research expertise and well-established domestic high-performance computing infrastructure, a better option may be to leverage those to procure and specialize their own AI compute – just as Canada is doing.
Organize a national community of AI practitioners through AI institutes
Making the best use of AI compute requires a national community of users that are connected, aware of the capacity, and informed about access. Under the CIFAR Pan-Canadian AI Strategy, the government supports three affiliated AI institutes: Vector in Toronto, Mila in Montreal and Amii in Edmonton. Each is full of world-renowned researchers that connect with AI practitioners in university programs, the private sector, hospital research departments and machine-learning associations in the community at large. The institutes touch most areas in Canada where AI learning, application and general tinkering take place. They are hubs that can connect a nationwide AI community, and this puts them in a perfect position to organize, support and communicate with AI practitioners.
Let AI practitioners guide AI compute allocation
AI progress is fast and dynamic, and practitioners can identify valuable ways to use AI compute that may only be apparent to those closely following the field. For instance, with open-source tools, the latest optimizations and access to AI compute that supports quick experimentation with large complex datasets, researchers can unlock new ways to train models with greater efficiency, thereby increasing compute availability and research productivity. Doing so depends on giving a prominent role in decisions about how AI compute should be used to experts who deeply understand the AI landscape.
Allot AI compute to projects of societal and economic importance
AI compute should also be used for more than just fundamental research. Health care research, AI commercialization demonstrations and Responsible AI – a field including fairness, explainability and reliability – are areas that have important impacts on society and deserve attention. AI researchers can use national compute capacity to advance these priorities as well.
The world is turning attention to the need for more AI compute. How countries expand capacity will influence how effectively new compute is used and whether it supports strategic goals. Here, Canada has set up an approach that’s promising and worth calling attention to. It’s also one that other countries can look to emulate as they address their own capacity constraints.
Your time is valuable. Have the Top Business Headlines newsletter conveniently delivered to your inbox in the morning or evening. Sign up today.