In-memory computing on a photonic platform (1801.06228v1)

Abstract: Collocated data processing and storage are the norm in biological systems. Indeed, the von Neumann computing architecture, that physically and temporally separates processing and memory, was born more of pragmatism based on available technology. As our ability to create better hardware improves, new computational paradigms are being explored. Integrated photonic circuits are regarded as an attractive solution for on-chip computing using only light, leveraging the increased speed and bandwidth potential of working in the optical domain, and importantly, removing the need for time and energy sapping electro-optical conversions. Here we show that we can combine the emerging area of integrated optics with collocated data storage and processing to enable all-photonic in-memory computations. By employing non-volatile photonic elements based on the phase-change material, Ge2Sb2Te5, we are able to achieve direct scalar multiplication on single devices. Featuring a novel single-shot Write/Erase and a drift-free process, such elements can multiply two scalar numbers by mapping their values to the energy of an input pulse and to the transmittance of the device, codified in the crystallographic state of the element. The output pulse, carrying the information of the light-matter interaction, is the result of the computation. Our all-optical approach is novel, easy to fabricate and operate, and sets the stage for development of entirely photonic computers.

• The paper reports an innovative integration of phase-change materials enabling all-photonic in-memory scalar multiplication with 13 distinct transmission levels.
• The authors employ optimized pulse widths of 25 ns and a 2.5 MHz Write/Erase cycle to deliver high-speed and low-energy computation.
• The study’s approach offers promising implications for machine learning and optical communications by mitigating data transfer bottlenecks and reducing energy consumption.

An Evaluation of In-Memory Computing on a Photonic Platform

The paper "In-memory computing on a photonic platform" presents a significant advancement in the field of optical computing by integrating data storage and processing within a single photonic device. The authors explore the inherent potential of integrated photonic circuits for on-chip computation by employing phase-change materials (PCMs) to execute in-memory computing tasks. This paper reports a successful implementation of all-photonic scalar multiplication using phase-change memories, with promising implications for high-speed, low-energy computational tasks.

Key Contributions and Methodology

The authors present an innovative approach that leverages non-volatile photonic elements, primarily phase-change materials like Ge$_2$Sb$_2$Te$_5$ (GST), to facilitate collocated data storage and processing on a single device. This work builds on the concepts of memcomputing but transitions the paradigm into the optical domain, which can lead to significant improvements in speed and reduction of energy consumption.

The paper demonstrates the direct multiplication of scalar numbers in the range [0,1] using single integrated photonic phase-change memory cells. The researchers achieve this by mapping the values of the scalar numbers to the energy of an input light pulse and the transmittance of the photonic device, respectively. The resultant pulse, which carries the information from the light-matter interaction, signifies the computational outcome.

Results

The paper reports the development of a drift-free, efficient mechanism for scalar multiplication that outperforms previous methods relying on sequential addition in computational efficiency. The use of multilevel memory states, established through varying pulse energies, enables 13 distinct transmission levels within the GST memory cell.

The implemented photonic device demonstrates promising energy efficiency and speed. The authors optimized pulse widths to 25 ns for amorphization and used a novel single-step pulse for erasure, achieving significantsavingsoverearlierapproaches. This results in an operation frequency of 2.5 MHz for a Write/Erase cycle, with a total switching energy of approximately 577 pJ for erasure.

Additionally, the material exhibits notable stability over long durations, up to 10 seconds, without detectable drift when a continuous probe is used. This lack of drift, when compared to electronic analogs, enhances reliability for sustained computation.

Implications and Future Directions

The implications of this work extend both practically and theoretically. Practically, such photonic devices could transform machine learning applications by affording efficient matrix-vector multiplication, a cornerstone operation in machine learning tasks. The integration of computational capabilities with memory on photonic chips potentially enables new architectures in optical communications and massive data processing, mitigating the traditional bottlenecks associated with electronic data transitions.

Theoretically, this approach could foster further investigation into photonic computing paradigms, particularly concerning mixed-precision workloads where photonic devices can be combined with traditional electronic elements to balance precision and performance.

In conclusion, this research contributes a robust foundation for exploring optical computing's potential, indicating a pathway towards refining and expanding upon existing computational paradigms with integrated phase-change photonic devices. Future work could explore the scalability of this approach in more complex computational tasks and hybrid integrations with other computational systems and architectures.