Skip to main content

Photonic-assisted Neuromorphic Computing system

Periodic Reporting for period 1 - PHOTON-NeuroCom (Photonic-assisted Neuromorphic Computing system)

Reporting period: 2017-08-01 to 2019-07-31

Conventional computing systems (like CPUs in PCs, Smartphones and etc.) will inevitably within the next 10 years reach a limit because of fundamental scientific reasons including limits on manufacturing, speed, density, transistor technology , power constraints . The most promising solution to this are the brain-inspired computing systems, so-called neuromorphic computing system (NCS). Implementation of NCSs using conventional transistor technology (CMOS) is area- and power-inefficient. Such inefficiencies have driven a significant effort to investigate the development of beyond-CMOS NCSs. The non-CMOS implementation of synapse has been researched to be implemented by spin-based materials (memristors, MTJs, STNOs and etc.). Despite some progress, still there is a huge difference (5-6 orders of magnitude) between the performance (operation/sec/Watt/cm3) of state-of-the-art NCS and human brain.
Neuromorphic computing market is expected to reach USD 1.7 billion by 2025 with an annual growth of 86%. Such growth shows the importance of neuromorphic computing and it is expected to find a huge market with the exponential growth of data processing, especially images and videos. In this respect, the social impact of PHOTON-NeuroCom’s technology can be tremendous. PHOTON-NeuroCom’s technology would allow for on-device computation, saving energy, speed, and improving privacy. PHOTON-NeuroCom will bring huge impact on society and EU economy, and the possibilities of new market creation is overwhelming, e.g. medical devices, edge devices, drones, space applications, military, IoT, smart vehicles, surveillance, smart cameras, financial forecasting, data mining, life-long self-learning machines etc.
The overall aim of PHOTON-NeuroCom was to realize a novel integration platform that combines photonic with current combination of the spin-based material and electronic (i.e. spintronic) in order to achieve an energy-efficient and high-speed brain-inspired computing system. The overall objectives of PHOTON-NeuroCom can be listed as:
1-Modelling the effect of heating on the dynamic and static behaviors of MTJ/STNO and the interaction between laser and MTJ/STNO
2-Design and simulation of a real-time laser-assisted MTJ/STNO-based NCS
3-Validating the models and designed systems by experimental results
The envisioned photonic-assisted neuromorphic computing system in PHOTON-NeuroCom is shown in Figure 1. In PHOTON-NeuroCom, for the first time, to the best of our knowledge, the proof-of-the-concept of thermally assisting a STNO-based NCS through a laser pulse was explored. Moreover, the effect of assisting an MTJ-based NCS through laser pulses was studied as an extra work. To do this, the following steps have been taken:
1- The effect of the laser on the temperature of the MTJ/STNO is explored through the COMSOL simulations and a behavioral model of MTJ/STNO interaction has been extracted.
2- A Verilog model of MTJ with embedded thermal effect on its dynamic and static behaviors is developed from an already available LLG-based model (Fong et al.; 2014). Then, in order to validate the model, it is benchmarked to experimental results of (Takeuchi et al.;2015).
3- A robust track and terminate circuit for real-time NCS is designed and its effect on the higher performance of real-time NCS has been shown.
4- Based on the above mentioned models, the first photonic-assisted MTJ-based NCS is designed and simulated. The designed system showed significantly higher performance in terms of energy consumption, speed and energy-delay product compared with the typical NCS.
5- The effect of heating on the performance of STNOs has been investigated through experimental results during two weeks stay at international Iberian Nanotechnology Laboratory - INL. Based on the achieved experimental results, the first laser assisted STNO-based NCS has been designed and its higher performance compared with typical NCSs is shown.
In general the achieved results can be divided to two sections:
1- Laser assisted MTJ-based NCS
Delay: The mean value (μDelay) and the standard deviation (σDelay) of the delay are calculated as shown in Figure 3A. Increasing the temperature from 27°C to 127°C reduces the 6-sigma worst case of MTJ switching delay by 84%.
Energy consumption: The total energy consumption of the laser assisted MTJ-based NCS includes the energies in the spintronic layer (memristors and MTJs), the CMOS interfacing circuit and the laser. Figure 3B shows the energy consumption of an MTJ and the memristors connected the MTJ to the inputs in spintronic layer at different temperatures. The total energy consumption of MTJ switching decreases by 80% while increasing the temperature to 127°C. Accordingly, the energy consumption of sensing circuit and firing the post NCS are calculated to be around 107fJ. The average energy consumption reduction of a column of memristors connecting their corresponding MTJ to inputs is 90% by increasing the temperature to 127°C (Figure 3B). The energy consumption per stimulation phase of the CMOS interface circuit is shown in Figure 3C at different temperatures. The energy consumption of the laser source is shown in Figure 3C. The energy consumption of laser per process (stimulation and recovery phase) is estimated as 29.4fJ using COMSOL simulation. The total energy consumption of the laser assisted MTJ-based NCS decreases by 85.7% at 127°C compared with 27°C, see Figure 3C.
Energy-delay product: Finally, the energy-delay product (EDP) of the whole laser assisted MTJ-based NCS is calculated (Figure 3D). The EDP of the laser assisted MTJ-based NCS is 1.34ns×pJ at 127°C, which is improved by 97.8% compared with the typical NCS at 27°C (60.7ns×pJ).
2- Laser assisted STNO-based NCS
Experimental results: Figure 4 shows the experimental setup used for characterizing the STNO samples. Figure 5A shows the PSD measured at different temperature from 27°C to 100°C for 230μA bias current (the curves are offset by 10μV2 along the vertical axis for clarity). Figure 5B shows the measured STNO resistance in P- and AP-state. The TMR ratio decreases by increasing the temperature (Figure 5C). The matched output power (Pout) of the STNO versus the bias current is shown in Figure 5D. In order to eliminate the effect of noise on Pout, the output power of the negative bias currents are deducted from the output power of the positive bias currents as shown in Figure 5E. Elevating the temperature increases Pout as shown in Figure 5D,E.
Simulation results: The effectiveness of the proposed laser assisted STNO-based NCS is evaluated by the hand-written digit recognition application. A 196×10 NCS is designed to recognize the handwritten digits in MATLAB. In order to model the effect of temperature increase on the power consumption of the STNOs, equations are fitted to the experimental results of Figure 5 and used in MATLAB. As illustrated in Figure 6B,C, increasing the temperature from 27°C to 100 has reduced the average power consumption of the memristor array and the STNOs by 54.7% and 55.3%, respectively. Hence, the total power consumption of the spintronic layer is reduced by 54.9% (Figure 6D). The power consumption of the CMOS interfacing circuit and the laser are shown in Figure 6E. In recovery step, the laser is illuminated for 2ns with 213μW power to keep the temperature of the STNOs around 100°C. The total power consumption of the laser assisted STNO-based NCS decreases by 40%.
The proposed real-time CMOS readout circuit.
The envisioned photonic-assisted neuromorphic computing system.
Simulation results of the laser assisted MTJ-based NCS
The STNO microscopic image, the schematic view of its stack and the experimental setup used
Experimental results
Simulation results laser assisted STNO-based NCS at hand-written digit recognition application