GPU Auxiliary Surface Access Across Discontiguous Memory Blocks

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In graphics processing units with distributed memory systems, compressed data is stored in discontiguous physical addresses, leading to bandwidth bottlenecks during storage and retrieval of auxiliary surfaces, and inefficient retrieval of auxiliary data for multiple discontiguous blocks of compressed data.

Innovation Solution

Assigning discontiguous blocks of compressed data to match hash calculations, reducing the number of memory accesses and improving memory throughput by minimizing redundant data retrieval.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If compressed data is stored in discontiguous physical addresses in a distributed memory system, then data compression is achieved, but bandwidth bottlenecks occur during storage and retrieval of auxiliary surfaces

Engineering Contradiction:

Improvedata compression ratioVSAvoidmemory bandwidth throughput

Core Design Contradiction:

Quantity of substanceVSProductivity

Solution Approach 1:

The patent merges the retrieval operations for auxiliary surfaces associated with multiple discontiguous blocks of compressed data into a single memory access operation. By combining these previously separate access operations, the system retrieves all necessary auxiliary data in one go, eliminating the bandwidth bottlenecks caused by multiple sequential accesses to discontiguous memory locations.

Inventive Principle:

Principle #5Merging (Combining)

2Reliability

If auxiliary data is retrieved for multiple discontiguous blocks of compressed data, then complete data retrieval is achieved, but the number of memory accesses increases causing inefficiency

Engineering Contradiction:

Improvedata retrieval completenessVSAvoidmemory access time

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The patent combines multiple separate memory access operations into a single unified access operation. Instead of accessing memory separately for each discontiguous block of compressed data, the system performs one consolidated access that retrieves all necessary auxiliary surfaces, thereby maintaining complete data retrieval while significantly reducing the time lost to multiple sequential memory accesses.

Inventive Principle:

Principle #5Merging (Combining)

3Quantity of substance

If auxiliary surfaces are stored and retrieved in a distributed memory system, then memory capacity is increased, but bandwidth bottlenecks are created during access

Engineering Contradiction:

Improvememory capacityVSAvoidaccess throughput

Core Design Contradiction:

Quantity of substanceVSProductivity

Solution Approach 1:

The patent addresses the bandwidth bottleneck by merging multiple auxiliary surface access operations into a single memory access. This approach maintains the increased memory capacity benefits of distributed memory systems while improving access throughput by eliminating the sequential access pattern that caused the bottleneck in the first place.

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentUS20250335356A1Auxiliary compression control surface access in a multi-node memory system

Publication Date: 2025.10.30 INTEL CORP

AI summary

A system that includes a graphics processing unit (GPU) comprising multiple processors and circuitry to: store compressed first data into the at least one memory based on a first address; store meta data associated with the compressed first data into the at least one memory; store compressed second data into the at least one memory based on a second address; copy the meta data into a cache of the at least one cache; and decompress the compressed first data and compressed second data based on at least one read of the stored meta data from the cache of the at least one cache.