Tech giant you’ve probably never heard of wants to put a data center in a shoebox build using 1000 3D superconducting chips — at 20 exaflops, it would be 20x faster than the most powerful supercomputer on Earth
To counter the strain on global energy resources due to accelerating computational demands – yes, AI, we’re looking at you – research institute imec is suggesting a radical shift away from traditional computing methods.
Its solution, detaile in IEEE Spectrum engineering magazine, involves exploiting the fundamental properties of superconductors to greatly reduce energy consumption, thereby creating an innovative, superconducting processor.
This promising technology has been in development for a couple of years so far and uses standard CMOS fabrication techniques which can potentially offer computing power that is a hundred times more energy-efficient than today’s best chips, and which could “lead to a computer that fits a data-center’s worth of computing resources into a system the size of a shoebox.”
20 exaflops
Imec’s research involved designing a new kind of processor from the ground up, with close collaboration between CMOS engineers and full-stack development teams.
Imec switched from using niobium as it superconducting material to the related compound niobium titanium nitride as it can withstand temperatures used in CMOS fabrication without losing its superconducting capabilities.
The resultant superconducting chip, optimized for AI processors, resembles a typical 3D CMOS system-on-chip. However, one significant difference is that most of the chip must be submerged in liquid helium to ensure optimal operating temperature, close to 4 Kelvin.
In comparison to conventional CMOS chips, superconductors dissipate only a fraction of the energy in the form of heat. This feature enables the possibility of stacking computational chips directly on top of each other, reducing the physical footprint while preserving the density gains brought about by Moore’s Law.
Imec estimates that a stack of 100 superconducting boards, each with superconductor processing units (SPUs), superconducting SRAM, and DRAM memory stacks and housed in a 20 x 20 x 12 centimeter shoebox-sized cooling environment, can perform a staggering 20 exaflops in BF16 number format. This is 20x the capability of the most powerful supercomputer currently in existence (Frontier at the Oak Ridge National Laboratory, Tennessee) consuming a total of 500 kilowatts of power. It also offers 100x the level of energy efficiency.
Imec’s innovation does not aim to replace existing CMOS computing technologies, but rather to supplement them. Institute researchers believe the tech will boost AI and machine learning, and integrate seamlessly with quantum computers.
The smaller data center footprint made possible by superconducting digital technology will allow these systems to be placed closer to their target applications, ultimately yielding colossal advancements in industries like agriculture, healthcare, and energy.
More from TechRadar Pro
To counter the strain on global energy resources due to accelerating computational demands – yes, AI, we’re looking at you – research institute imec is suggesting a radical shift away from traditional computing methods. Its solution, detaile in IEEE Spectrum engineering magazine, involves exploiting the fundamental properties of superconductors to…
