Accelerator Interface Unit for Domain Specific Accelerators

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Integrating domain specific accelerators (DSAs) with processing systems onto the same chip is a complex task due to the need for substantial changes in the instruction set architecture (ISA) and toolchain, requiring a simplified approach to facilitate efficient acceleration and processing.

Innovation Solution

A novel approach involving minor modifications to the toolchain, where a processing system includes a main processor, an accelerator interface unit with interface registers, and domain specific accelerators, allowing for efficient communication and execution of commands between the main processor and accelerators, with new instructions added to the ISA to manage DSA operations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If domain specific accelerators are integrated with processing systems onto the same chip, then computation scalability and processing efficiency are improved, but device complexity and toolchain modification requirements increase substantially

Engineering Contradiction:
Improvecomputation scalabilityVSAvoidtoolchain modification requirements
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

An accelerator interface unit is introduced as an intermediary component between the main processor and domain specific accelerators. This interface unit manages communication protocols, handles command routing, and coordinates data transfer, thereby simplifying the integration process and reducing the complexity of toolchain modifications required while enabling scalable computation architecture

Inventive Principle:
Principle #24Intermediary (Mediator)

2Productivity

If domain specific accelerators are integrated with processing systems onto the same chip, then processing efficiency is improved, but interaction latency between processor and accelerator increases

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidinteraction latency
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The accelerator interface unit is integrated within the same chip as both the main processor and domain specific accelerators, merging these components into a unified architecture. This integration enables direct communication paths and shared memory spaces, significantly reducing interaction latency while maintaining high processing efficiency through coordinated operation of the interface unit and accelerators

Inventive Principle:
Principle #5Merging (Combining)

3Adaptability or versatility

If instruction set architecture is modified to accommodate DSA operations, then adaptability and programming capability are improved, but ease of manufacture and toolchain development deteriorate

Engineering Contradiction:
Improveprogramming capabilityVSAvoidtoolchain development
Core Design Contradiction:
Adaptability or versatilityVSEase of manufacture

Solution Approach 1:

The accelerator interface unit implements a universal communication interface that supports multiple domain specific accelerators through standardized protocols. This multi-functional interface unit handles various accelerator types and operation modes, improving adaptability and programming capability while simplifying toolchain development through consistent, standardized interaction patterns rather than requiring custom interfaces for each accelerator type

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

Data Source

PatentUS20230393851A1Processing system with integrated domain specific accelerators
Publication Date: 2023.12.07 ALIBABA GROUP HOLDING LTD
  • US20230393851A1 patent drawing
  • US20230393851A1 patent drawing
  • US20230393851A1 patent drawing

AI summary

A number of domain specific accelerators (DSA1-DSAn) are integrated into a conventional processing system (100) to operate on the same chip by adding additional instructions to a conventional instruction set architecture (ISA), and further adding an accelerator interface unit (130) to the processing system (100) to respond to the additional instructions and interact with the DSAs.