20240012871. ITERATION ENGINE FOR THE COMPUTATION OF LARGE KERNELS IN CONVOLUTIONAL ACCELERATORS simplified abstract (STMICROELECTRONICS S.r.l.)

ITERATION ENGINE FOR THE COMPUTATION OF LARGE KERNELS IN CONVOLUTIONAL ACCELERATORS

Organization Name

STMICROELECTRONICS S.r.l.

Inventor(s)

Antonio De Vita of Milano (IT)

Thomas Boesch of Rovio (CH)

Giuseppe Desoli of San Fermo Della Battaglia (IT)

ITERATION ENGINE FOR THE COMPUTATION OF LARGE KERNELS IN CONVOLUTIONAL ACCELERATORS - A simplified explanation of the abstract

This abstract first appeared for US patent application 20240012871 titled 'ITERATION ENGINE FOR THE COMPUTATION OF LARGE KERNELS IN CONVOLUTIONAL ACCELERATORS

Simplified Explanation

The abstract describes a convolutional accelerator that performs convolution operations on a streaming feature data tensor using a kernel. It includes a feature line buffer, a kernel buffer, a multiply-accumulate cluster, and iteration control circuitry.

• The convolutional accelerator convolves a kernel with a streaming feature data tensor.
• The kernel is decomposed into multiple sub-kernels.
• The sub-kernels are iteratively convolved with respective sub-tensors of the streamed feature data tensor.
• The iteration control circuitry defines windows of the streamed feature data tensors corresponding to the sub-tensors.

Potential Applications:

• Image and video processing: The convolutional accelerator can be used for tasks like object detection, image recognition, and video analysis.
• Deep learning: It can accelerate the computation of convolutional neural networks (CNNs) used in various applications such as natural language processing and computer vision.

Problems Solved:

• Efficient convolution: The accelerator optimizes the convolution process by decomposing the kernel and convolving sub-kernels with sub-tensors, reducing computational complexity.
• Real-time processing: The streaming feature data tensor allows for real-time processing of data, enabling applications that require low latency.

Benefits:

• Faster processing: The accelerator improves the speed of convolution operations, enabling quicker analysis and decision-making.
• Resource efficiency: By decomposing the kernel and using sub-tensors, the accelerator reduces memory and computational requirements.
• Real-time capabilities: The streaming feature data tensor and iteration control circuitry enable real-time processing, making it suitable for time-sensitive applications.

Original Abstract Submitted

a convolutional accelerator includes a feature line buffer, a kernel buffer, a multiply-accumulate cluster, and iteration control circuitry. the convolutional accelerator, in operation, convolves a kernel with a streaming feature data tensor. the convolving includes decomposing the kernel into a plurality of sub-kernels and iteratively convolving the sub-kernels with respective sub-tensors of the streamed feature data tensor. the iteration control circuitry, in operation, defines respective windows of the streamed feature data tensors, the windows corresponding to the sub-tensors.