Intermediate Representation for AI Chip Data Migration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional technologies lack an intermediate representation (IR) construction method for migrating data between multiple storage locations in AI chips with domain-specific architecture (DSA), which complicates the execution of operations on tensors across various storage locations.
Innovation Solution
An intermediate representation construction method that generates IRs including data migration information and segmentation details, allowing tensors to be migrated through multiple storage locations, with the ability to express migration paths and tensor sizes, enabling efficient data movement between on-chip buffers or memory in AI chips.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If data is migrated through multiple storage locations in DSA-based AI chips, then computing capability and operator complexity increase, but the complexity of managing data migration paths and storage locations increases
Solution Approach 1:
The patent segments the data migration process into discrete paths between storage locations, with each path independently managed. The intermediate representation divides the tensor migration into segments that can be tracked and optimized separately, reducing the complexity of managing multi-location data movement.
Solution Approach 2:
The patent introduces an intermediate representation (IR) as a mediator between the high-level computing operations and the physical data migration paths. This IR layer abstracts the complexity of multiple storage locations and migration paths, providing a simplified interface for managing data movement while enabling sophisticated computing capabilities.
2Loss of energy
If multiple storage locations are used to shorten data movement distance, then data movement overheads are reduced, but the complexity of constructing and managing intermediate representations increases
Solution Approach 1:
The patent performs preliminary construction of the intermediate representation before executing data migration operations. By pre-computing and optimizing the IR structure with knowledge of the storage location topology, the system can efficiently manage data movement without incurring high construction complexity during execution.
Solution Approach 2:
The patent utilizes parameter transformations within the intermediate representation to optimize data migration paths. By changing parameters such as tensor segmentation and migration timing in the IR, the system achieves efficient data movement while keeping the construction process manageable through systematic parameter optimization.
3Productivity
If tensor migration paths are optimized, then data movement efficiency improves, but the complexity of determining optimal paths through multiple storage locations increases
Solution Approach 1:
The patent creates a simplified copy or abstraction of the storage location topology within the intermediate representation. This copied model allows the system to determine optimal migration paths through simulation and analysis without directly managing the physical complexity of multiple storage locations and interconnections.
Data Source
AI summary
An intermediate representation construction method is provided. The method includes: obtaining a first intermediate representation (IR), where the first IR includes a computing statement, the computing statement includes a tensor and an operator, an operation represented by the operator is performed by a computing unit, the computing unit is configured to perform an operation on data that is migrated through a first storage location and a second storage location sequentially, and the tensor is data that is used when the operation represented by the operator is performed; and generating a second IR based on the computing statement, where the second IR includes first data migration information and data segmentation information.