54 lines
3.3 KiB
Markdown
54 lines
3.3 KiB
Markdown
#### **Heterogeneous-Accelerator Migration Sub-group**
|
|
|
|
* **Project Facts**
|
|
**Project Creation Date:** 2023/08/08
|
|
**Primary Contact:** Lei Huang & huangleiyjy@chinamobile.com
|
|
**Project Lead:** Lei Huang & huangleiyjy@chinamobile.com
|
|
**Committers:**? Lei Huang & huangleiyjy@chinamobile.com
|
|
**Mailing List:**[https://computing-force-network@lists.opendev.org](%3Cspan%3E%3Cdiv%3Ehttps://computing-force-network@lists.opendev.org%3C/div%3E%3C/span%3E)
|
|
**Meetings:** Same as CFN bi-weekly meeting
|
|
**Repository:** [https://opendev.org/cfn/computing-native](%3Cspan%3E%3Cdiv%3Ehttps://opendev.org/cfn/computing-native%3C/div%3E%3C/span%3E)
|
|
**StoryBoard:** [https://storyboard.openstack.org/#!/project_group/computing-force-network](%3Cspan%3E%3Cdiv%3Ehttps://storyboard.openstack.org/#!/project_group/computing-force-network%3C/div%3E%3C/span%3E)
|
|
|
|
* **Introduction**
|
|
|
|
The intelligent computing ecology is mainly composed of
|
|
middleware/framework + tool chain + hardware. Each vendors build
|
|
corresponding tool chain around its own hardware, and generates branch
|
|
version matching different AI framework.
|
|
|
|
The ecosystem is becoming diverse, cross-architecture and cross-stack
|
|
migration of upper-layer applications is extremely complex, which brings
|
|
development challenges to application developers, computing force
|
|
service providers and chip vendors.
|
|
|
|
In order to facing ecological challenges, we proposed a technology named
|
|
Heterogeneous-Accelerator Migration Technology , the goal of which is to
|
|
break the existing compile-link-execute tightly coupled tool chain
|
|
ecology, establish a new collaboration mechanism.Shield the underlying
|
|
hardware differences, and realize cross-architecture non-sensing
|
|
migration and execution of applications. Build traction model of the
|
|
intelligent computing industry chain with software as the core, and
|
|
prosper the ecology of intelligent computing industry.
|
|
|
|
The Heterogeneous-Accelerator Migration Technology architecture mainly
|
|
consists of two layers: the heterogeneous-accelerator migration
|
|
abstraction layer and the computing force pooling layer. Among them, the
|
|
heterogeneous-accelerator migration abstraction layer mainly includes
|
|
native interfaces based on a unified programming model and converters,
|
|
as well as a hardware-native stack formed by a cross-architecture
|
|
comprehensive compiler and runtime, generating a unified executable
|
|
program format. The computing force pooling layer mainly consists of
|
|
components for heterogeneous computing power registration management,
|
|
scheduling, and pooling, achieving unified management and pooled
|
|
execution of heterogeneous computing power resources.
|
|
|
|
|
|
Fig. Initial Architecture of Heterogeneous-Accelerator Migration Technology
|
|
|
|
* **Release Planning & Release Notes**
|
|
**1. Heterogeneous-Accelerator Migration Technology Solution**
|
|
Introduction of Heterogeneous-Accelerator Migration Technology solution, including user guide, architecture description, etc.
|
|
**2.Key component image of Heterogeneous-Accelerator Migration Platform**
|
|
The model calculation graph and program code for user input will be compiled and run across architectures, and key tool components will be provided.
|