Processing circuit for fusing computations in neural network
Patent Information
- Application Number
- TW115117829
- Authority / Receiving Office
- TW · TW
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2026-03-06
- Filing Date
- 2026-05-07
- Publication Date
- 2026-11-11
- Estimated Expiration
- 2046-05-06
Smart Images

Figure TWG2TB001921152_001 
Figure TWG2TB001921152_002 
Figure TWG2TB001921152_003
Abstract
Claims
1. A processing circuit, comprising: a vector processing engine and a matrix processing engine; A data routing circuit coupled to a memory, the matrix processing engine and the vector processing engine, and comprising: a first multiplexer having: a first input terminal coupled to the memory; A second input terminal is coupled to the matrix processing engine; And an output terminal, coupled to the vector processing engine; A second multiplexer has: a first input terminal coupled to the memory; A second input terminal is coupled to the vector processing engine; And an output terminal, coupled to the matrix processing engine; And a third multiplexer having: a first input coupled to the vector processing engine; A second input terminal is coupled to the matrix processing engine; And an output terminal, coupled to the memory; And a data routing controller, coupled to the data routing circuit.
2. The processing circuitry as described in claim 1, wherein the vector processing engine is used to perform a normalization layer operation, and the matrix processing engine is used to perform a matrix multiplication operation or a convolution operation.
3. The processing circuit as described in claim 2, wherein the operation of the normalization layer is decomposed into a simplification suboperation and a normalization suboperation.
4. The processing circuit as described in claim 3, wherein the simplification suboperation is fused with a first matrix or convolution operation, and the normalization suboperation is fused with a second matrix or convolution operation.
5. The processing circuitry as described in claim 1, wherein the first multiplexer controls the vector processing engine to select to receive data from the memory or the matrix processing engine, the second multiplexer controls the matrix processing engine to select to receive data from the memory or the vector processing engine, and the third multiplexer controls the memory to select to receive data from the vector processing engine or the matrix processing engine.
6. The processing circuit as described in claim 1, wherein in a first operation flow, the second multiplexer transfers first data from the memory to a first matrix processing unit of the matrix processing engine, the first multiplexer transfers a first operation result from the first matrix processing unit to a first vector processing unit of the vector processing engine, and the third multiplexer transfers a second operation result from the first vector processing unit to the memory.
7. The processing circuit as described in claim 6, wherein the first matrix processing unit performs a first matrix multiplication or convolution operation on the first data to produce the first operation result, and the first vector processing unit performs a simplification suboperation on the first operation result to produce the second operation result.
8. The processing circuit as described in claim 1, wherein in a second operation flow, the first multiplexer transfers a second piece of data from the memory to a second vector processing unit of the vector processing engine, the second multiplexer transfers a third operation result of the second vector processing unit to a second matrix processing unit of the matrix processing engine, and the third multiplexer transfers a fourth operation result of the second matrix processing unit to the memory.
9. The processing circuit as described in claim 8, wherein the second vector processing unit performs a normalization suboperation on the second data to produce the third operation result, and the second matrix processing unit performs a second matrix multiplication or convolution operation on the third operation result to produce the fourth operation result.
10. The processing circuitry as described in claim 1, wherein a first piece of data read from the memory is first processed by the matrix processing engine and then by the vector processing engine, and a second piece of data read from the memory is first processed by the vector processing engine and then by the matrix processing engine.
Citation Information
Patent Citations
Hardware acceleration circuit, data processing acceleration method, chip and accelerator
CN116306825A
Object recognition system based on tactile sensing array and neural network algorithm
CN117251760A
Artificial neural network processing methods and systems
EP4589482A1
Accumulator for digital computation-in-memory architectures
TW202324213A
