Dynamic Deep Learning Processor Runtime Reconfiguration

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deep learning processor architectures are statically customized and fail to efficiently handle dynamic variations in deep learning algorithms and applications, such as mixed sparse and dense layers and varying numerical precision, leading to inefficient performance.

Innovation Solution

A dynamically reconfigurable deep learning processor is implemented using a programmable logic device with partial reconfiguration capabilities, driven by self-monitoring and optimization systems that adjust numerical precision and architecture based on runtime data and performance metrics to match dynamic demands.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If a statically customized deep learning processor architecture is used, then the processor can be optimized for specific algorithms, but it cannot efficiently handle dynamic variations in deep learning algorithms and applications

Engineering Contradiction:
Improveadaptability to varying deep learning algorithmsVSAvoidprocessor architecture complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent implements a dynamically reconfigurable deep learning processor that can change its architecture at runtime based on the specific algorithm being executed. The processor includes reconfigurable processing elements, programmable interconnect structures, and dynamic memory allocation that allow the system to adapt its computational fabric to match the requirements of different deep learning workloads, transitioning from static to dynamic architecture.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent creates a universal deep learning processor platform that can handle multiple types of algorithms including convolutional neural networks, recurrent neural networks, and transformer models. The reconfigurable architecture provides multi-functionality by allowing the same hardware platform to be dynamically adapted to execute various algorithm types without requiring separate dedicated hardware for each algorithm.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Productivity

If the processor architecture is fixed, then manufacturing is simpler, but performance cannot be optimized for mixed sparse and dense layers with varying numerical precision

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidarchitecture implementation complexity
Core Design Contradiction:
ProductivityVSEase of manufacture

Solution Approach 1:

The patent employs parameter changes by allowing dynamic adjustment of numerical precision (e.g., switching between 8-bit, 16-bit, and 32-bit operations), activation function selections, and layer configuration parameters at runtime. The processor can reconfigure its operational parameters to match the specific requirements of sparse vs. dense layers, optimizing processing efficiency for each layer type without requiring different hardware designs.

Inventive Principle:
Principle #35Parameter changes

3Adaptability or versatility

If a programmable logic device is used to provide flexibility, then the device can adapt to various system demands, but it may not fully utilize the flexibility for static deep learning processor implementations

Engineering Contradiction:
Improveflexibility to adjust to system demandsVSAvoiddeep learning algorithm execution efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent implements preliminary action by pre-configuring the processor with common deep learning operation templates and dataflow patterns. Before executing a specific algorithm, the system pre-loads appropriate configuration parameters, activates relevant processing elements, and prepares memory structures in advance, allowing the programmable logic device to quickly adapt to system demands without full reconfiguration overhead.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4075290A1Dynamic deep learning processor architecture
Publication Date: 2022.10.19 ALTERA CORP
  • EP4075290A1 patent drawingFigure 1~2
  • EP4075290A1 patent drawingFigure 3
  • EP4075290A1 patent drawingFigure 4

AI summary

Methods and systems for dynamically reconfiguring a deep learning processor by operating the deep learning processor using a first configuration. The deep learning processor then tracking one or more parameters of a deep learning program executed using the deep learning processor in the first configuration. The deep learning processor then reconfigures the deep learning processor to a second configuration to enhance efficiency of the deep learning processor executing the deep learning program based at least in part on the one or more parameters.