To combat the strain that accelerating computing demands (yes, AI) are putting on the world's energy resources, research institute imec is proposing a radical shift away from traditional computing methods.
The solution, detailed in IEEE Spectrum engineering magazine, leverages the fundamental properties of superconductors to dramatically reduce energy consumption and create revolutionary superconducting processors.
This promising technology has been in development for several years now, uses standard CMOS manufacturing techniques, and could potentially deliver computing power that is 100 times more energy efficient than today's best chips, potentially leading to “a computer with the computing resources of a data center packed into a system the size of a shoebox.”
20 exaflops
In Imec's research, a new kind of processor has been designed from the ground up, in close collaboration between CMOS engineers and a full-stack development team.
Instead of using niobium as the superconducting material, Imec switched to using niobium titanium nitride, a related compound that can withstand the temperatures used in CMOS manufacturing without losing its superconducting function.
The resulting superconducting chips optimized for AI processors are similar to typical 3D CMOS systems-on-chips, with one key difference: To ensure optimal operating temperatures (close to 4 Kelvin), most of the chip must be immersed in liquid helium.
Compared to conventional CMOS chips, superconductors dissipate only a fraction of their energy in the form of heat, a feature that makes it possible to stack computational chips directly on top of one another, reducing their physical footprint while maintaining the density gains brought about by Moore's Law.
Imec estimates that a stack of 100 superconducting boards, each with a superconducting processing unit (SPU), superconducting SRAM, and DRAM memory stacks, housed in a cooled environment the size of a shoebox measuring 20 x 20 x 12 centimeters, could run an astounding 20 exaflops in the BF16 numerical format — 20 times the power of the most powerful supercomputer in existence today (Frontier at Tennessee's Oak Ridge National Laboratory), which consumes a total of 500 kilowatts of power — and be 100 times more energy efficient.
Imec's innovation is intended to complement, rather than replace, existing CMOS computing technology, and researchers at the institute believe the technology will advance AI and machine learning and seamlessly integrate with quantum computers.
Superconducting digital technology is shrinking the footprint of data centers, allowing these systems to be placed closer to their target applications, ultimately enabling breakthroughs in industries such as agriculture, healthcare and energy.
