Configurable Spatial Accelerator Dataflow Execution

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional processor architectures face challenges in achieving exascale performance and energy efficiency, particularly in executing dataflow graphs, as they struggle with out-of-order scheduling, complex register files, and high energy consumption.

Innovation Solution

A spatial array of processing elements connected by lightweight, back-pressured communication networks, where each processing element and network dataflow endpoint circuit performs operations only when input data is available and storage space is ready, eliminating control overheads and utilizing dataflow operators to execute dataflow graphs directly.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If conventional processor architectures are used to execute dataflow graphs, then general-purpose computing capability is maintained, but control overhead increases and energy consumption rises

Engineering Contradiction:

Improvecomputing capabilityVSAvoidcontrol overhead

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system segments the processing architecture into specialized processing elements arranged in a spatial array, each capable of executing dataflow operators. This segmentation eliminates the need for complex centralized control mechanisms while maintaining versatile computing capability through the distributed nature of the array.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

Processing elements in the spatial array are designed to autonomously execute operations based on available input data and storage readiness, eliminating control overhead. The elements self-regulate their execution based on dataflow availability rather than requiring external scheduling control.

Inventive Principle:

Principle #25Self-service

2Productivity

If conventional processor architectures with out-of-order scheduling are used, then instruction throughput is improved, but energy consumption and architectural complexity increase

Engineering Contradiction:

Improveinstruction throughputVSAvoidenergy consumption

Core Design Contradiction:

ProductivityVSUse of energy by moving object

Solution Approach 1:

The dataflow execution model allows processing elements to automatically execute operations when inputs are available and storage is ready, eliminating the need for complex out-of-order scheduling mechanisms. This self-service approach maintains high throughput while significantly reducing energy consumption by removing unnecessary control logic.

Inventive Principle:

Principle #25Self-service

3Adaptability or versatility

If conventional processor architectures with complex register files are used, then scheduling flexibility is improved, but energy consumption and hardware complexity increase

Engineering Contradiction:

Improvescheduling flexibilityVSAvoidregister file complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The invention extracts and removes complex register files and scheduling mechanisms from the architecture. Instead, it uses a simpler spatial array of processing elements that rely on dataflow availability and storage readiness to determine execution, thereby reducing hardware complexity while maintaining scheduling flexibility through the dataflow model.

Inventive Principle:

Principle #2Taking out (Extraction)

4Use of energy by moving object

If spatial array of processing elements is used to execute dataflow graphs, then energy efficiency and computational density are improved, but control mechanisms are simplified

Engineering Contradiction:

Improveenergy efficiencyVSAvoidcontrol mechanisms

Core Design Contradiction:

Use of energy by moving objectVSDevice complexity

Solution Approach 1:

Processing elements in the spatial array autonomously determine when to execute operations based on input data availability and storage readiness, eliminating the need for complex external control mechanisms. This self-service approach achieves high energy efficiency and computational density while using minimal control overhead.

Inventive Principle:

Principle #25Self-service

Data Source

PatentUS11593295B2Apparatuses, methods, and systems for operations in a configurable spatial accelerator

Publication Date: 2023.02.28 INTEL CORP

AI summary

Systems, methods, and apparatuses relating to operations in a configurable spatial accelerator are described. In one embodiment, a configurable spatial accelerator includes a first processing element that includes a configuration register within the first processing element to store a configuration value that causes the first processing element to perform an operation according to the configuration value, a plurality of input queues, an input controller to control enqueue and dequeue of values into the plurality of input queues according to the configuration value, a plurality of output queues, and an output controller to control enqueue and dequeue of values into the plurality of output queues according to the configuration value.