Monolithic multi-bit weight cell for neuromorphic computing
A neuromorphic weight cell (NWC) including a resistor ladder including a plurality of resistors connected in series, and a plurality of shunting nonvolatile memory (NVM) elements, each of the shunting NVM elements being coupled in parallel to a corresponding one of the resistors.
1. A neuromorphic weight cell (NWC) comprising:
a resistor ladder comprising a plurality of resistors connected in series; and
a plurality of shunting nonvolatile memory (NVM) elements, each of the shunting NVM elements being coupled in parallel to a corresponding one of the resistors.
2. The NWC of claim 1 , wherein a first resistor of the resistors at a top of the resistor ladder has a lowest resistance value of the resistors, and
wherein resistance values of each subsequent one of the resistors is twice a resistance value of an immediately preceding one of the resistors.
3. The NWC of claim 2 , wherein each of the shunting NVM elements has a first state of resistance and a second state of resistance, the first state of resistance being greater than the resistance value of the first resistor, and the second state of resistance being less than the resistance value of the first resistor.
4. The NWC of claim 1 , wherein terminals of each of the shunting NVM elements are individually addressable.
5. The NWC of claim 1 , wherein each of the shunting NVM elements comprise a two-terminal memristor, wherein the resistors comprise a vertical stack of doped poly layers.
6. An artificial neural network comprising a plurality of layers, each layer comprising a plurality of neural nodes,
wherein an input of one of the neural nodes is configured to be adjustably weighted by a neuromorphic weight cell (NWC) implemented in hardware, the NWC comprising:
a resistor ladder comprising a plurality of resistors connected in series; and
a plurality of shunting nonvolatile memory (NVM) elements, each of the shunting NVM elements being coupled in parallel to a corresponding one of the resistors,
wherein the input of the one of the neural nodes is configured to be adjustably weighted by selectively operating one or more of the shunting NVM elements.
7. The artificial neural network of claim 6 , wherein the shunting NVM elements each comprise a flash transistor.
8. The artificial neural network of claim 7 , wherein gates of each of the shunting NVM elements are connected to a common gate terminal.
9. The artificial neural network of claim 8 , wherein the NWC is configured to be read as a weight by:
floating all source and drain connections of the flash transistors; and
applying a common voltage to the common gate terminal.
10. The NWC of claim 9 , wherein a cell weight of the NWC is configured to be programmed by individually programming the flash transistors,
wherein a programmed flash transistor of the flash transistors comprises a source terminal set to ground and a drain terminal set to a high voltage,
wherein all source and drain terminals of the flash transistors above the programmed flash transistor are set to ground,
wherein all source and drain terminals of the flash transistors below the programmed flash transistor are set to the high voltage, and
wherein gates of all of the flash transistors are set to the high voltage.
11. The NWC of claim 8 , wherein a cell weight of the NWC is configured to be erased by grounding all source and drain terminals of all of the flash transistors, and setting the common gate terminal to a tunneling voltage, or by setting the common gate terminal to ground, and setting all of the source and drain terminals of all of the flash transistors to a tunneling voltage.
12. The NWC of claim 8 , wherein the flash transistors are implemented as a vertical stack on a common poly-Si channel, and wherein the resistors are implemented as a separate vertical stack.
13. The NWC of claim 12 , wherein resistance values of the resistors are set by in-situ doping, and wherein a doping concentrations of the resistors vary as respective factors of two.
14. The NWC of claim 13 , wherein sources and drains of the flash transistors comprise heavily-doped Si layers, and are coupled with respective contacts of the resistors, and
wherein a doping concentration of the heavily-doped Si layers is at least an order of magnitude higher than a doping concentration of a most heavily doped one of the resistors.
15. The NWC of claim 14 , wherein the heavily-doped Si layers are individually contacted by trench contacts.
16. The NWC of claim 15 , wherein a number of layers of the flash transistors is greater than or equal to two.
17. The NWC of claim 16 , wherein the NWC is configured to be read by floating all of the sources and drains of the flash transistors, and by applying a read voltage to an input terminal of the NWC.
18. A method of forming a neuromorphic weight cell (NWC) comprising a resistor ladder comprising a plurality of resistors connected in series, and a plurality of shunting nonvolatile memory (NVM) elements, each of the shunting NVM elements being coupled in parallel to a corresponding one of the resistors, the method comprising:
depositing a common polycrystalline Si channel on an insulating substrate;
depositing a first spacer material layer on the common polycrystalline Si channel;
depositing an insulating material as a sacrificial gate on the first spacer material layer;
depositing a second spacer material layer on the insulating material;
forming a plurality of flash transistors as a first vertical stack on the common polycrystalline Si channel; and
forming a plurality of resistors as a second vertical stack on the common polycrystalline Si channel.
19. The method of claim 18 , further comprising selectively in-situ doping areas of the second vertical stack to set different resistance values of the resistors.
20. The method of claim 18 , further comprising performing etching and deposition to form trench contacts for individually contacting source and drain contacts of the flash transistors.