Matrix-vector multiplication using SOT-based non-volatile memory cells
An apparatus is provided that includes an array including n rows and m columns of nodes, each row of nodes coupled to one of n first conductive lines, each column of nodes coupled to one of m second conductive lines, each node of the n rows and m columns of nodes including a spin orbit torque MRAM non-volatile memory cell configured to store a corresponding weight of an n×m array of weights each having a first weight value or a second weight value, and a control circuit configured to apply n input voltages each having a first input value or a second input value to corresponding n first conductive lines, the n input voltages corresponding to an n-element input vector. The spin orbit torque MRAM non-volatile memory cells are configured to generate m output currents at the m second conductive lines upon application of the n input voltages. The m output currents corresponding to a result of multiplying the input vector by the n×m array of weights.
1. An apparatus comprising:
an array comprising n rows and m columns of nodes, each row of nodes coupled to one of n first conductive lines, each column of nodes coupled to one of m second conductive lines, each node of the n rows and m columns of nodes comprising a spin orbit torque magnetoresistive memory (MRAM) non-volatile memory cell configured to store a corresponding weight of an n×m array of weights each having a first weight value or a second weight value; and
a control circuit configured to apply n input voltages each having a first input value or a second input value to corresponding n first conductive lines, the n input voltages corresponding to an n-element input vector, wherein:
each spin orbit torque MRAM non-volatile memory cell comprises:
a first terminal coupled to one of the n first conductive lines;
a second terminal coupled to one of the m second conductive lines; and
a third terminal coupled to one of n third conductive lines;
the spin orbit torque MRAM non-volatile memory cells are configured to generate m output currents at the m second conductive lines upon application of the n input voltages; and
the m output currents correspond to a result of multiplying the input vector by the n×m array of weights.
2. The apparatus of claim 1 , wherein:
each spin orbit torque MRAM non-volatile memory cell further comprises:
a Spin Hall Effect layer; and
a magnetic tunnel junction comprising a pinned layer and a free layer that each have a perpendicular direction of magnetization; and
the magnetic tunnel junction comprises a substantially cylindrical shape.
3. The apparatus of claim 1 , wherein:
each spin orbit torque MRAM non-volatile memory cell further comprises:
a Spin Hall Effect layer; and
a magnetic tunnel junction comprising a pinned layer and a free layer that each have an in-plane direction of magnetization;
the magnetic tunnel junction comprises a substantially ellipsoidal shape; and
the pinned layer and the free layer each have an easy axis perpendicular to a current flow in the Spin Hall Effect layer.
4. The apparatus of claim 1 , wherein:
each spin orbit torque MRAM non-volatile memory cell further comprises:
a Spin Hall Effect layer; and
a magnetic tunnel junction comprising a pinned layer and a free layer that each have an in-plane direction of magnetization;
the magnetic tunnel junction comprises a substantially ellipsoidal shape; and
the pinned layer and the free layer each have an easy axis that is at an angle off-perpendicular to a current flow in the Spin Hall Effect layer.
5. The apparatus of claim 1 , wherein:
each spin orbit torque MRAM non-volatile memory cell further comprises:
a Spin Hall Effect layer; and
a magnetic tunnel junction comprising a free layer; and
each spin orbit torque MRAM non-volatile memory cell is configured to be programmed by:
applying a first electrical current pulse through the Spin Hall Effect layer to cause a direction of magnetization of the free layer to oscillate; and
applying a second electrical pulse through the magnetic tunnel junction and the Spin Hall Effect layer to switch the direction of magnetization of the free layer.
6. The apparatus of claim 1 , wherein each spin orbit torque MRAM non-volatile memory cell further comprises a Spin Hall Effect layer comprising any of a heavy metal and a topological insulator.
7. The apparatus of claim 1 , wherein each spin orbit torque MRAM non-volatile memory cell further comprises a Spin Hall Effect layer comprising bismuth antimony with (012) orientation.
8. The apparatus of claim 1 , wherein:
the n input voltages correspond to elements of an input vector; and
each spin orbit torque MRAM non-volatile memory cell is configured to conduct a memory cell current that corresponds to a result of multiplying one of the elements of the input vector by the corresponding weight stored in the spin orbit torque MRAM non-volatile memory cell.
9. The apparatus of claim 1 , wherein:
the n input voltages correspond to elements of an input vector; and
each of the m output currents corresponds to a sum of products of the input vector with corresponding weights in a column of the n×m array of weights.
10. An apparatus comprising:
a cross-point memory array comprising a plurality of spin orbit torque magnetoresistive memory (MRAM) non-volatile memory cells configured to store synaptic weights of an artificial neural network;
a plurality of word lines coupled to the spin orbit torque MRAM non-volatile memory cells; and
a plurality of bit lines coupled to the spin orbit torque MRAM non-volatile memory cells, wherein:
each spin orbit torque MRAM non-volatile memory cell comprises a Spin Hall Effect layer comprising a topological insulator; and
in response to a plurality of input voltages coupled to the plurality of word lines, the spin orbit torque MRAM non-volatile memory cells are configured to generate output currents at the plurality of bit lines representing outputs of the artificial neural network.
11. The apparatus of claim 10 , wherein each spin orbit torque MRAM non-volatile memory cell further comprises:
a first terminal coupled to one word line of the plurality of word lines; and
a second terminal coupled to one bit line of the plurality of bit lines.
12. The apparatus of claim 10 , wherein:
each spin orbit torque MRAM non-volatile memory cell further comprises a magnetic tunnel junction comprising a free layer; and
each spin orbit torque MRAM non-volatile memory cell is configured to be programmed by:
applying a first electrical current pulse through the Spin Hall Effect layer to cause a direction of magnetization of the free layer to oscillate; and
applying a second electrical pulse through the magnetic tunnel junction and the Spin Hall Effect layer to switch the direction of magnetization of the free layer.
13. The apparatus of claim 10 , wherein:
each spin orbit torque MRAM non-volatile memory cell further comprises a magnetic tunnel junction comprising a free layer; and
each spin orbit torque MRAM non-volatile memory cell is configured to be programmed by applying an electrical current pulse through the Spin Hall Effect layer to switch a direction of magnetization of the free layer.
14. The apparatus of claim 10 , wherein each bit line of the plurality of bit lines is input to an activation function of the artificial neural network.
15. The apparatus of claim 10 , wherein the topological insulator is bismuth antimony with (012) orientation.
16. A method comprising:
programming each cell of a plurality of spin orbit torque magnetoresistive memory (MRAM) non-volatile memory cells with a corresponding weight of an n×m array of weights, each cell having a first weight value or a second weight value, wherein:
each spin orbit torque MRAM non-volatile memory cell comprises:
a Spin Hall Effect layer; and
a magnetic tunnel junction comprising a free layer; and
each spin orbit torque MRAM non-volatile memory cell is configured to be programmed by:
applying a first electrical current pulse through the Spin Hall Effect layer to cause a direction of magnetization of the free layer to oscillate; and
applying a second electrical pulse through the magnetic tunnel junction and the Spin Hall Effect layer to switch the direction of magnetization of the free layer;
generating n input voltages corresponding to an n-element input vector; and
applying the n input voltages to the plurality of spin orbit torque MRAM non-volatile memory cells to generate m output currents that represent a result of multiplying the n-element input vector by the n×m array of weights.
17. The method of claim 16 , wherein the Spin Hall Effect layer comprises a topological insulator.
18. The method of claim 16 , further comprising, in response to a plurality of input voltages coupled to a plurality of word lines coupled to the spin orbit torque MRAM non-volatile memory cells:
generating, by the spin orbit torque MRAM non-volatile memory cells, output currents at a plurality of bit lines coupled to the spin orbit torque MRAM non-volatile memory cells and representing outputs of an artificial neural network, wherein the plurality of spin orbit torque MRAM non-volatile memory cells are configured to store synaptic weights of the artificial neural network.
19. The apparatus of claim 1 , wherein each spin orbit torque MRAM non-volatile memory cell further comprises a Spin Hall Effect layer comprising a topological insulator.
20. The apparatus of claim 1 , further comprising means for training the n×m array of weights to generate the m output currents representing outputs of an artificial neural network.