Three dimensional circuit implementing machine trained network
Some embodiments provide a three-dimensional (3D) circuit structure that has two or more vertically stacked bonded layers with a machine-trained network on at least one bonded layer. As described above, each bonded layer can be an IC die or an IC wafer in some embodiments with different embodiments encompassing different combinations of wafers and dies for the different bonded layers. The machine-trained network in some embodiments includes several stages of machine-trained processing nodes with routing fabric that supplies the outputs of earlier stage nodes to drive the inputs of later stage nodes. In some embodiments, the machine-trained network is a neural network and the processing nodes are neurons of the neural network. In some embodiments, one or more parameters associated with each processing node (e.g., each neuron) is defined through machine-trained processes that define the values of these parameters in order to allow the machine-trained network (e.g., neural network) to perform particular operations (e.g., face recognition, voice recognition, etc.). For example, in some embodiments, the machine-trained parameters are weight values that are used to aggregate (e.g., to sum) several output values of several earlier stage processing nodes to produce an input value for a later stage processing node.
1. A three-dimensional (3D) circuit comprising:
a first bonded layer; and
a second bonded layer vertically stacked with the first bonded layer;
wherein the first bonded layer comprises a machined-trained network and a first memory for the machine-trained network, and the second bonded layer comprises a second memory for the machine-trained network;
machined-trained network comprising a plurality of machined-trained processing nodes;
the first memory to store machine-trained parameters for the processing nodes, and the second memory to store values produced by machine-trained processing nodes.
2. The 3D circuit of claim 1 , wherein the machine-trained network is a neural network and the processing nodes are neurons.
3. The 3D circuit of claim 1 , wherein at least one bonded layer is an integrated circuit (IC) die and the other bonded layer is an IC wafer.
4. The 3D circuit of claim 1 , wherein the first and second bonded layers are first and second integrated circuit (IC) dies.
5. The 3D circuit of claim 4 further comprising a package that includes a case that encapsulates the first and second dies within one housing.
6. The 3D circuit of claim 1 , wherein the first and second bonded layers are first and second integrated circuit (IC) wafers.
7. The 3D circuit of claim 1 , wherein the second memory is connected to the machine-trained network through a vertical conductive interface.
8. The 3D circuit of claim 7 , wherein the vertical conductive interface comprises through-silicon vias (TSVs) or direct bonded interconnect (DBI) connections.
9. The 3D circuit of claim 7 , wherein the second memory includes a plurality of memory arrays and at least two of the memory arrays are connected to the machine-trained network through the vertical conductive interface.
10. The 3D circuit of claim 1 , wherein the first and second memories are different types of memories.
11. The 3D circuit of claim 10 , wherein the first memory is a non-random access memory, while the second memory is a random access memory (RAM).
12. The 3D circuit of claim 1 , wherein the first bonded layer comprises a third memory to store additional values produced by the machine-trained processing nodes, and the second bonded layer comprises a third memory to store additional machine-trained parameters for the processing nodes.
13. The 3D circuit of claim 1 , wherein each bonded layer comprises (i) an integrated circuit substrate on which a plurality of circuit components are defined and (ii) a plurality of wiring layers for connecting the circuit components.
14. A three-dimensional (3D) circuit comprising:
a first bonded layer; and
a second bonded layer vertically stacked with the first bonded layer;
wherein the first bonded layer comprises a machined-trained network and a first memory for the machine-trained network, and the second bonded layer comprises a second memory for the machine-trained network;
machined-trained network comprising a plurality of machined-trained processing nodes;
the first memory to store values produced by machine-trained processing nodes, and the second memory to store machine-trained parameters for the processing nodes.
15. The 3D circuit of claim 14 , wherein the second memory is connected to the machine-trained network through a vertical conductive interface that comprises through-silicon vias (TSVs) or direct bonded interconnect (DBI) connections.
16. The 3D circuit of claim 14 , wherein the first and second memories are different types of memories.
17. The 3D circuit of claim 16 , wherein the first memory is a sequentially accessed read-only memory (ROM), while the second memory is an ephemeral random access memory (RAM).
18. The 3D circuit of claim 16 , wherein the first memory comprises an output interface that sequentially reads storage locations in the first memory, while the second memory comprises an output interface that supports random access of storage locations of the second memory.
19. The 3D circuit of claim 14 , wherein the second bonded layer comprises a third memory to store additional values produced by the machine-trained processing nodes.
20. The 3D circuit of claim 14 , wherein the first bonded layer comprises a third memory to store additional machine-trained parameters for the processing nodes.
21. An integrated circuit (IC) comprising:
an arrangement of neurons forming a neural network;
a first memory to store machine-trained parameters for the neurons; and
a second memory to store values produced by the neurons, said second memory being separate from the first memory.
22. The IC of claim 21 , wherein the first and second memories are different types of memories.
23. The IC of claim 22 , wherein the first memory is a non-random access memory, while the second memory is a random access memory (RAM).
24. The IC of claim 23 , wherein the first memory is a sequentially accessed read only memory (ROM), while the second memory is a read-write memory.
25. The IC of claim 21 further comprising first and second vertically stacked dies, wherein the first die comprises the neural network while the second die comprises at least one of the first and second memories.
26. A three-dimensional (3D) circuit comprising:
a first bonded layer; and
a second bonded layer vertically stacked with the first bonded layer;
wherein the first bonded layer comprises a machine-trained network, and the second bonded layer comprises first and second memories for the machine-trained network;
the machined-trained network comprising a plurality of machined-trained processing nodes;
the first memory to store values produced by the processing nodes, and the second memory to store machine-trained parameters for the processing nodes.