GPU Cache Write-Back Scheduling During Frame Processing

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The bottleneck in bandwidth and performance caused by evicting dirty cache lines in the cache hierarchy of graphics processing units (GPUs) during frame transitions, due to the significant bandwidth and processing power required for writing back data to system memory, is not effectively addressed by existing technologies.

Innovation Solution

Concurrently writing back dirty cache lines to system memory while processing frames, with priority determination based on read command occupancy at the system memory controller, and marking cache lines as clean post-write to avoid redundant writes during frame transitions.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If dirty cache lines are written back to system memory during frame transitions, then data consistency is maintained, but bandwidth consumption increases and performance deteriorates

Engineering Contradiction:
Improvedata consistencyVSAvoidGPU performance
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent applies preliminary action by writing back dirty cache lines during the current frame processing rather than waiting for frame transition. The system proactively flushes dirty cache lines to memory while the graphics pipeline is still processing the current frame, before the frame ends and before the next frame begins. This prevents the bandwidth bottleneck that would occur during frame transitions when all dirty cache lines must be written back simultaneously.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If all dirty cache lines are written back during frame transition, then data integrity is ensured, but memory bandwidth is consumed and processing time increases

Engineering Contradiction:
Improvedata integrityVSAvoidframe transition time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The system performs cache line write-back operations during the current frame's execution time rather than deferring them to frame transition. By initiating the write-back process early while the frame is still being processed, the system completes data integrity operations before the frame ends, eliminating the time loss that would occur during frame transitions.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent maintains continuity of useful action by overlapping the write-back operation with the graphics pipeline processing. Instead of stopping frame processing to complete write-backs, the system continuously performs both operations in parallel, ensuring that data integrity is maintained without interrupting the productive work of the graphics pipeline.

Inventive Principle:
Principle #20Continuity of useful action

3Loss of energy

If cache lines are marked as clean after write-back, then redundant writes are avoided, but write-back timing must be precisely controlled

Engineering Contradiction:
Improvebandwidth consumptionVSAvoidcache management complexity
Core Design Contradiction:
Loss of energyVSDevice complexity

Solution Approach 1:

The patent implements feedback by setting a status bit in the cache line metadata to indicate when a cache line has been written back to memory. This feedback mechanism allows the cache controller to recognize which cache lines have been flushed and should not be written back again during frame transitions, preventing redundant write operations and reducing bandwidth consumption.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system changes the state parameter of cache lines from dirty to clean (or written-back) by setting status bits in the cache metadata. This parameter change allows the cache management hardware to distinguish between cache lines that need to be written back and those that have already been processed, enabling precise control of write-back timing without requiring complex software management.

Inventive Principle:
Principle #35Parameter changes

4Productivity

If write-back priority is based on read command occupancy, then bandwidth is optimized, but priority determination logic becomes more complex

Engineering Contradiction:
Improvememory bandwidth efficiencyVSAvoidpriority determination logic
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent applies self-service by having the cache controller automatically determine write-back priority based on the current system state (read command occupancy at the memory controller). The system monitors its own workload conditions and autonomously adjusts write-back timing without external intervention, optimizing bandwidth utilization while avoiding the need for complex external priority management logic.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4168898B1Selectively writing back dirty cache lines concurrently with processing
Publication Date: 2025.11.12 ADVANCED MICRO DEVICES INC
  • EP4168898B1 patent drawingFigure 1
  • EP4168898B1 patent drawingFigure 2
  • EP4168898B1 patent drawingFigure 3

AI summary

A graphics pipeline (200) includes a cache (207, 305) having cache lines (310-313) that are configured to store data used to process frames in the graphics pipeline. The graphics pipeline is implemented using a processor (115) that processes frames for the graphics pipeline using data stored in the cache. The processor processes a first frame and writes back a dirty cache line from the cache to a memory (105) concurrently with processing of the first frame. The dirty cache line is retained in the cache and marked as clean subsequent to being written back to the memory. In some cases, the processor generates a hint that indicates a priority for writing back the dirty cache line based on a read command occupancy at a system memory controller (335).