This artist's impression shows how an array of ring resonators covered with a magneto-optical layer could function as a matrix multiplier, acting on multiwavelength input vectors. [Image: Brian Long, Senior Artist, UCSB]
Researchers are working on a wide range of technologies to try and satisfy the ever-increasing appetite for processing power on the part of artificial intelligence and machine learning. Now, an international group of electrical engineers has shown that this demand might be met by photonic memory based on what is known as nonreciprocal magneto-optics (Nat. Photonics, doi: 10.1038/s41566-024-01549-1). Their system, which they have run for billions of cycles, is not only fast and efficient but also robust.
Moore’s law lags
Despite the decades-long march of Moore's law―which tells us that the number of transistors per integrated circuit doubles roughly once every two years―semiconductor manufacturers are struggling to keep up with the rise of new processor-hungry applications. According to the company Open AI, in just six years (from 2012 to 2018) the amount of computing power needed to train the largest deep-learning networks rose by more than a factor of 300,000, whereas graphics processing units became a mere 300 times more efficient over the same period.
Photonics could potentially help bridge this gap, thanks to light's greater speed and more efficient passage through integrated circuits. In particular, numerous groups are targeting so-called photonic in-memory computing, which carries out the matrix multiplications central to neural networks by using a matrix of programmable optical weights to linearly transform input vectors encoded in optical signals.
Scaling photonics
One of the main outstanding problems with such schemes is scalability. Whereas electronic in-memory processing systems known as crossbar arrays typically carry out an entire matrix multiplication in one pass, integrated photonic circuits can only store a tiny fraction of all the weights that make up the matrix. The solution is to reprogram the photonic weights many times for each matrix multiplication, but this places a premium on speed, efficiency and endurance.
Different groups have demonstrated various aspects of the technology, but no one has yet shown that they can simultaneously satisfy the multiple demanding requirements. For example, researchers have demonstrated that electronic memristors integrated into waveguides can carry out fast and efficient switching in a nonvolatile way (requiring power only to change, rather than maintain, states), but such devices are not robust, having failed to go beyond 1,000 cycles of writing and erasing data.
Harnessing magnetic fields
Although the researchers only tested single cells, they demonstrated the potential of the technology―updating the cells in just a single nanosecond and consuming just over one ten-thousandth of a nanojoule per bit.
In the latest work, Nathan Youngblood, University of Pittsburgh, USA; Paolo Pintus, University of California, Santa Barbara, USA, and the University of Cagliari, Italy; and colleagues in the United States and Japan have obtained impressive values of all the necessary parameters by using magnetic fields to control the phase of microring resonators covered with a magneto-optical layer.
Subject to an applied field, the resonators' two counter-propagating modes―one traveling clockwise and the other counterclockwise―experience phase shifts in opposite directions along the electromagnetic spectrum. The upshot is that incoming red-shifted laser light experiences far greater resonance in one direction than the other, which leads to a measurable difference in transmission from the two modes, encoding the value of that particular weight.
Many such cells placed in an array would carry out matrix multiplication on an input vector of multicolored laser pulses. Each pulse in each row would couple to a specific resonator (according to the latter's diameter), with the resonators' outputs then added together in a balanced photodetector. This would yield two values (the transmitted intensities of the two modes) that together would constitute one component of the array's output vector.
The researchers showed that they could make 70-µm diameter resonators by combining cerium-substituted yttrium iron garnet (Ce:YIG) with silicon in two different ways, both potentially compatible with CMOS manufacturing techniques. The first of these involves bonding a Ce:YIG wafer to a silicon waveguide, while the second instead sees an amorphous silicon layer deposited on the Ce:YIG and patterned. Both approaches use a gold electromagnet placed on top of the stack to set up a magnetic field.
A look ahead
Although the researchers only tested single cells, they demonstrated the potential of the technology―updating the cells in just a single nanosecond and consuming just over one ten-thousandth of a nanojoule per bit. Crucially, they did so for 2.4 billion write-and-erase cycles without any sign of wear, which is over a million times more cycles than the next-best technology for this parameter (a phase-change material). The team also showed that such operations could be made nonvolatile by incorporating patterned cobalt–iron–boron magnetic stripes into the resonators' cladding.
Challenges remain. In particular, says Youngblood, going from single cells to arrays will involve making smaller rings and less-lossy Ce:YIG. Looking slightly further ahead, he adds that it should be possible to boost cells' switching efficiency by exploiting spin-orbit-torque or spin-torque-transfer effects―since these would obviate the need for an integrated electromagnet.