This article is more than 1 year old
MIT, Amazon, TSMC, ASML and friends to work on non-planet-killing AI hardware
Program co-lead tells us 'energy efficiency is the greatest need'
Big names in tech are collaborating with academics to develop energy-optimized machine-learning and quantum-computing systems under the MIT AI Hardware Program, an initiative announced on Tuesday.
Chip makers like TSMC and Analog Devices, hardware development lab NTT Research, supplier of EUV machines ASML, and tech behemoth Amazon have signed up so far.
The goal is to figure out a roadmap outlining the production of next-generation, energy-efficient hardware for AI and quantum computing in the coming decade. To this end, the research will focus on developing novel architectures and software at the heart of a range of technologies, from analog neural networks and neuromorphic computing, to hybrid-cloud computing and HPC. Designs will be tested using proofs of concept at MIT.nano, the US university's small-scale fabrication facility.
The MIT AI Hardware Program will be co-led by Jesús del Alamo, professor of electrical engineering, and Aude Oliva, director of strategic industry engagement and the MIT-IBM Watson AI Lab. It will be chaired by Anantha Chandrakasan, dean of the School of Engineering and a professor of electrical engineering and computer science at the university.
"In the past few years, we have seen seemingly superhuman capabilities of AI systems," Alamo told The Register.
"Properly used, they are poised to transform many human activities such as transportation, health, education, defense, etc. As progress in algorithms and data sets continues at a brisk pace, hardware must keep up or the promise of AI will not be realized. That is why it is critically important that research takes place on AI hardware."
It's one thing to build more and more powerful chips and systems to handle ever-growing neural networks. It's another to do it in an energy-efficient manner that is sustainable for our planet, which is what this effort is focused on.
"More optimized hardware has been proposed but significant new research is needed to realize these concepts," Alamo said. "Energy efficiency is the greatest need. As data sets get bigger, the hardware needs to expand accordingly and the energy consumption just explodes. It does not scale. We need new hardware."
The MIT AI Hardware Program is being funded by industry, we're told. Alamo declined to say how much its inaugural members had chipped in so far.
- Google uses deep learning to design faster, smaller AI chips
- Nvidia's Morpheus AI security framework to land in April
- Cerebras brings wafer-size AI chips to medical data analysis
- If you want to make your own chip and aren't Microsoft rich, who do you turn to?
Supply chains for semiconductors are strained during the ongoing COVID-19 pandemic. Shortages in key materials and high demand has led to a backlog of component orders struggling to be cleared by foreign manufacturers, prompting governments around the world to invest in efforts ramping up homegrown chip production.
Intel CEO Pat Gelsinger just urged US Congress to pass the $52bn Creating Helpful Incentives to Produce Semiconductors (CHIPS) for America Act to subsidize factory expansions. Meanwhile, the European Commission has proposed €11bn (~$12.2bn) funding to bolster chip R&D under the European Chips Act.
Alamo said the MIT AI Hardware Program is focused on next-generation hardware for emergent technologies and is less concerned about the global chip crunch.
"The supply chain issues are transitory; we are thinking long term," he said. "University research is most effective five to ten years out and beyond." ®