Graphics Architecture Neural Pipeline for Adaptive GPU Prefetching

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional scheduling and dispatching in graphics hardware rely on prefetching subsequent consecutive cache lines, which can be improved using neural networks.

Innovation Solution

Implementing AI-based techniques such as AI-driven thread dispatch, AI-based dynamic scheduling, and AI-driven hardware memory prefetching to enhance the efficiency of GPU deep pipelines.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If conventional prefetching techniques are used to schedule and dispatch cache lines, then the scheduling process is simple and predictable, but the efficiency and adaptability of GPU deep pipelines are limited

Engineering Contradiction:

Improveefficiency of GPU deep pipelinesVSAvoidcomplexity of scheduling system

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent replaces conventional mechanical prefetching mechanisms with a neural network-based scheduling system. The neural network learns patterns from GPU workload traces and predicts optimal cache line prefetching decisions, substituting rigid hardware-based prefetching with adaptive AI-driven scheduling that can handle complex GPU deep pipeline workloads more efficiently

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The system implements self-service by having the neural network automatically learn from and adapt to workload patterns without manual programming. The neural network processes GPU workload traces, extracts features, trains models, and generates prefetching decisions autonomously, eliminating the need for hand-crafted prefetching algorithms and enabling the system to optimize itself based on observed behavior

Inventive Principle:

Principle #25Self-service

2Adaptability or versatility

If AI-based techniques are implemented to improve scheduling and prefetching, then the efficiency and adaptability of GPU deep pipelines are enhanced, but the complexity of the system increases

Engineering Contradiction:

Improveadaptability to different workloadsVSAvoidcomplexity of scheduling system

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by transforming raw GPU workload traces into extracted features through feature extraction layers. The neural network processes these features through multiple layers with different parameter configurations to learn workload patterns. This allows the system to adapt to different workload types by adjusting the feature extraction and neural network parameters rather than requiring complete system redesign

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20250252650A1Graphics architecture including a neural network pipeline

Publication Date: 2025.08.07 INTEL CORP

AI summary

One embodiment provides a graphics processor comprising a block of graphics cores and circuitry including a programmable neural network unit, the programmable neural network unit including one or more neural network hardware blocks, wherein a neural network hardware block includes circuitry to perform neural network operations and activation operations for a layer of a neural network, the programmable neural network unit addressable by cores within the block of graphics cores, wherein the programmable neural network unit is to configure one or more neural network hardware blocks with a meta-shader neural network, the meta-shader neural network to generate a texture for one of multiple types of terrain.