Magnetic tunnel junction beyond memory from logic to neuromorphic computing WANJUN PARK DEPT. OF ELECTRONIC ENGINEERING, HANYANG UNIVERSITY

Transcription

1 Magnetic tunnel junction beyond memory from logic to neuromorphic computing WANJUN PARK DEPT. OF ELECTRONIC ENGINEERING, HANYANG UNIVERSITY

2 Magnetic Tunnel Junctions (MTJs) Structure High density memory Free layer (FM 1) Tunneling barrier (Insulator: MgO) Pinned layer (FM 2) Bit line (BL) Word line (WL) AP or P state 0 or 1 Function Representation of non-volatile binary state according to magnetization configuration Source line (SL) 1 Transistor + 1 MTJ Advantage Scalability Low energy High speed High endurance CMOS compatibility 1GB STT-MRAM, Everspin (2016) - 2 -

3 Content To find computing functions beyond memory from MTJ for integrated circuits Construction of 2-input MTJ MTJ Logic gates Neuromorphic computing Artificial MTJ neuron Artificial MTJ synapse Artificial Neurotransmission system - 3 -

4 Basic configuration of MTJ Single input configuration for switching to achieve the binary state FM 1 FM 2 MTJ switching FM 1 FM 2 Anti-parallel state (R AP ) Parallel state (R P ) Methods for magnetization reversal of free layer Switching variable 1: magnetic field switching (current induced Ampere field) I Switching variable 2: Spin-transfer torque (STT) switching (spin polarized current) I - 4 -

5 Two-input configuration of MTJ Motivation for multiple input extension Increase of functional flexibility - Reduction of switching stress by breakup of biases - Increase of switching bias margin Physical variables for MTJ switching H in by applying I AF V in for I STT R MTJ - Ampere field-induce switching - Spin-transfer torque (STT) switching - Thermally assisted switching (TAS) I OF - Voltage-assisted switching (ME effect) - Spin-orbit torque (SOT) switching Multiple input is available for MTJ I STT X I AF Structure of 2-input MTJ Our choice: STT & Ampere field for two switching inputs Sharing integration methods developed for MRAM - 5 -

6 Switching characteristics of 2-input MTJ MTJs for switching characteristics due to mixed inputs of STT & Ampere field Size : (80 80) ~ ( ) nm 2-6 -

7 Interpretation of 2-input switching H a = 0 (at coercive center) H a = 30 Oe H a = -30 Oe Magnetic field assisted STT switching Case for switching to be P state Energy due to Ampere field (H ext = H a ) : Required energy for STT switching E b+ = E b0 - E b * H s : Switching field E 0 : Energy barrier at zero magnetic field a J : Spin transfer torque a c : Critical spin transfer torque β = 2-7 -

8 NAND/NOR representation Definition of binary states for each input Assignment of input values to the STT input terminal Logic value Input 1 (V) Input 2 (H) MTJ OUT 1 V H H H R AP 0 V L H L R P V H R V L =0.2 V(<V S1 ), V H =0.3 V (V S1 <V H <V S2 ) H L =-5 (Oe), H H = 5 (Oe) V H R V L =0.3 V (V S1 <V L <V S2 ), V H =0.4 V (>V S2 ) H L =-5 (Oe), H H = 5 (Oe) - 8 -

9 All logical representation founded in MTJ 7 Boolean logic representations from possible 12 binary inputs of voltage biases for the STT switching No. Initial Input set Input (V, H), Output (R) Logic function State V L V H H L H H (V L,H L ) (V H,H L ) (V L,H H ) (V H,H H ) R AP = 1, R P = 0 R AP = 0, R P = R AP R AP R AP R P NAND AND R AP R P R P R P NOR OR R R AP R P R AP R P NOT V V 4 AP R AP R AP R AP R AP TRUE FALSE R AP R AP R P R P NOT H H R 5 Oe 5 Oe P R P R P R P FALSE TRUE R P R AP R P R P V NIMP H V IMP H R AP R AP R P R AP H IMP V H NIMP V R R P R AP R P R AP V NIMP H V IMP H 10 P R P R P R P R P FALSE TRUE R AP R AP R P R P NOT H H R AP R AP R AP R AP TRUE FALSE - 9 -

10 MTJ Logic gate Logic gate for digital computing - Cascading computing Full schematic of a logic gate

11 14 Boolean functions computed in MTJ logic gate - Each function is confirmed by SPICE simulation modified with MTJ micro-model - XOR and XNOR are missing among full 16 Boolean logics

12 XOR/XNOR MTJ Logic gate - XOR/XNOR gate could be completed by using cascading computing Conclusively, we have two types of MTJ logic gate which allow any digital computing

13 Reconfigurable Logic Reconfigurability: further advantage of MTJ logic < Truth table >

14 Practical example for reconfigurable Logic Carry-out function with reconfigurable logic < Truth table >

15 Neurotransmission spike signal carrying information < Slowly Adapting> < Fast Adapting > Stimulus for tactile sense Pressure Frequency Slowly Adapting Fast Adapting

16 Neural coding Neural (Biological) coding To find carrier for information according to Strength and frequency of input stimulus - Rate coding Information spiking rate - Temporal coding Information Spiking pattern (Spike ordering in timing) - Time-to-first-spike, Phase Correlations, Spiking sequence, Synchrony etc

17 Telegraphic switching Telegraphic switching by mixed effect of STT & Ampere field in 2-inpt MTJ I = -10 µa μa H = 80 Oe R (kω) μa 10 μa 20 μa Time (s) Total Energy (E M + E STT ) ~ E S Toggling between AP and P state Stochastic characteristic Switching probability P(H, I) is defined as the carried information M. Pufall et al., Phys. Rev. B (2004)

18 Rate coding Construction of neural coding : Rate coding RRRRRRRR(BBBBBBBBBBBBBB) = nn ssss (nnnnnnnnnnnn oooo ssssssssssss TT (tttttttt wwwwwwwwwwww) RRRRRRRR MMMMMM = ) tt AAAA (oooo tt PP ) TT (tttttttttt bbbbbb oooo P AP 1 ( H, I) = 1+ exp( α H + βi + γ ) Information carrier : rate Stimulus : H and I applied by independent inputs

19 Artificial neuron function MTJ-based neuron architecture representing the rate coding

20 Neurotransmission Neuron system connected by Synapse (Spike-rate-dependent plasticity) Neuron # : 10 11~12 Synapse # : 10 14~ ~4 /neuron - Spike(=Action potential) generation (when the signal is above threshold) - Neural coding Information is coded through spike train Rate coding - Synaptic weight: plasticity for connection strength - Weight modulation Potentiation / Depression Spike-timing-dependent plasticity (STDP)

21 Memristive character of MTJ

22 MTJ-based artificial synapse Construction of input signal for Spike-timing dependent plasticity <Biological STDP> Δt > 0 Δt < 0 R. Froemke et al., Nature (2002)

23 Artificial neurotransmission Neurotransmission system Artificial neurotransmission system MTJs are commonly used for neural and synaptic functions Learning rule of Spike-rate dependent plasticity is possibly applied

24 SRDP learning rule

25 Summary MTJ was modified with two inputs for switching to achieve functional flexibility. Then we found various computing functions for digital to neuromorphic computing