Granted Patent B2

US 12,585,931 · App. 18/143,502 · Granted Mar 24, 2026

Inventor: Hsiang-Lan Lung (Ardsley, NY)

Assignee: MACRONIX INTERNATIONAL CO., LTD.

G06N3/063G06N3/048

Abstract

Claims (32)

1 . An artificial intelligence (AI) inference platform, comprising:

a logic die including an array of AI accelerator processing elements, each AI accelerator processing element including an activation memory and an AI processor element;

a memory die including an array of cells implementing a weight memory and a page buffer; and

a plurality of vertical connections that directly connect AI accelerator processing elements in the logic die and the page buffer of corresponding cells in the array of the memory die.

2 . The AI inference platform of claim 1 , wherein the AI accelerator processing elements implement neural processing units (NPU).

3 . The AI inference platform of claim 1 , wherein the direct vertical connections comprise copper-to-copper hybrid bond vias.

4 . The AI inference platform of claim 1 , wherein an AI processor element performs at least one of a storage and a retrieval of a neural network weight to or from a particular page buffer of a corresponding memory cell of the array for use in neural network computations conducted by a corresponding AI accelerator processing element in the logic die.

5 . The AI inference platform of claim 1 , wherein the memory die further comprises 3D NAND.

6 . The AI inference platform of claim 1 , wherein the activation memory comprises a static random-access memory (SRAM).

7 . The AI inference platform of claim 6 , wherein the activation memory stores activation data including a non-linear function identifying relationships between inputs and outputs of a neural network.

8 . The AI inference platform of claim 7 , wherein the activation memory stores activation data including data implementing an activation function selected from at least one of a sigmoid function, a hyperbolic tangent (tanh) function, a rectified linear unit (ReLU) function, a leaky rectified linear unit (LReLU) function, and a maxout function.

9 . The AI inference platform of claim 1 , wherein the memory die implements one of a phase change memory (PCM), a NOR flash memory, a 3D AND flash memory, a resistive random-access memory (RRAM), a magneto-resistive random-access memory (MRAM), a ferroelectric random-access memory (FeRAM), a conductive bridge random-access memory (CBRAM), and a 3D NAND flash memory.

10 . The AI inference platform of claim 1 , wherein the page buffer of a corresponding cell in the array of the memory die is incorporated into the memory die, such that no external buffer is located between the memory die and the logic die.

11 . An artificial intelligence (AI) inference method for applying weight data from a high-bandwidth memory to neural network computations, the method comprising:

retrieving, by an AI processor element of a plurality of AI processor elements arrayed on a logic die, a plurality of neural network weights stored in a corresponding 3D memory cell of a plurality of 3D memory cells arrayed on a memory die; wherein the AI processor element is directly connected to the corresponding 3D memory cell by a plurality of metal-to-metal hybrid bonds formed between metal contacts placed upon contacting surfaces of each of the logic die and the memory die;

applying, using an accelerator core coupled to the AI processor element, the plurality of neural network weights to input data for one or more nodes of a plurality of nodes of a neural network to obtain an intermediate output;

retrieving activation data from an activation memory directly connected with the AI processor element on the logic die;

applying the retrieved activation data to the intermediate output to obtain a result for a neural network layer; and

storing the result in at least one of the corresponding 3D memory cell, a field programmable gate array (FPGA) buffer, and the activation memory, thereby facilitating computations at additional neural network layers until a final result is reached.

12 . The artificial intelligence (AI) inference method of claim 11 , wherein the AI processor element is directly connected to the activation memory by a second plurality of metal-to-metal hybrid bonds formed between metal contacts placed upon contacting surfaces of each of the logic die and the memory die.

13 . The artificial intelligence (AI) inference method of claim 11 , wherein the retrieving of the activation data from the activation memory, further includes retrieving data representing a non-linear function identifying relationships between inputs and outputs of the neural network.

14 . The artificial intelligence (AI) inference method of claim 13 , wherein the retrieving of the activation data from the activation memory, further includes retrieving data implementing an activation function selected from at least one of a sigmoid function, a hyperbolic tangent (tanh) function, a rectified linear unit (ReLU) function, a leaky rectified linear unit (LReLU) function, and a maxout function.

15 . The artificial intelligence (AI) inference method of claim 11 , further including storing, to the activation memory, data representing a non-linear function identifying relationships between inputs and outputs of the neural network.

16 . The artificial intelligence (AI) inference method of claim 15 , wherein the storing of the data further includes storing data implementing an activation function selected from at least one of a sigmoid function, a hyperbolic tangent (tanh) function, a rectified linear unit (ReLU) function, a leaky rectified linear unit (LReLU) function, and a maxout function.

17 . The artificial intelligence (AI) inference method of claim 11 , further including:

receiving, by the AI processing element, weight data; and

storing the weight data as received to a memory cell of the weight memory directly connected to the AI processing element.

18 . The artificial intelligence (AI) inference method of claim 11 , further including:

receiving, by at least two AI processing elements of the plurality of AI processing elements, weight data; and

storing, by each of the at least two AI processing elements, the weight data as received to corresponding 3D memory cells of the at least two of the plurality of 3D memory cells directly connected to the at least two AI processing elements.

19 . The artificial intelligence (AI) inference method of claim 11 , further including:

repeatedly performing the retrieving of the neural network weights, the applying of the neural network weights, the retrieving of the activation data, the applying of the activation data and the storing of the result for additional neural network layers, until a final result is reached.

Assignments (1)

Continuity (1)

References Cited (369)