Table 2 Design requirements, main advantages, and key challenges of various computing schemes using integrated CMOS and MTJ

Design requirements

● In-memory logic: balancing nonvolatility and energy efficiency

● Neuromorphic computing:

 ◦ Large CMOS read signal for analog states;

 ◦ Efficient design by balancing the advantages of analog states and cost of CMOS analog to digital conversion

● Neuromorphic computing:

 ◦ Large read margin for RF output without need of or with minimal CMOS amplification circuits;

 ◦ Effective tuning of the oscillatory output using CMOS

● Efficient CMOS sampling (reading) of stochastic signals

● Efficient design of connections with CMOS

● Effective tuning of the stochastic state using CMOS

● Efficient CMOS sampling (reading) of chaotic signals

● Efficient design of connections with CMOS

● Effective tuning of the chaotic state using CMOS

Main advantages

● Direct integration with CMOS process

● Mature

● Versatile

● Intrinsic capability of handling RF signals (MHz-GHz)

● Phase, amplitude, and frequency encoding

● Tunable coupling

● Low power consumption by leveraging thermal fluctuations

● Tunable coupling

● Rich and tunable dynamics in one or few devices

Key challenges

● Multiple states with large readout margin

● Reducing I_C/Δ (power-stability tradeoff)

● Better RF interconnect

● System/algorithm development

● Multiple states with large readout margin

● Overcome device variation

● System/algorithm development

● Overcome device variation

● More experimental demonstration

● System/algorithm development

● Overcome device variation