Peer-to-Peer GPU Data Transfer via Fabric Interface Translation

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In modern computing environments, graphics processing units (GPUs) face inefficiencies due to the inability to read data from peer GPUs' frame buffers, leading to duplicated effort and increased overhead when using system memory for data transfers, as existing interconnectivity fabric standards like AGP and PCI-Express do not permit peer-to-peer reading.

Innovation Solution

A method and apparatus that enable peer-to-peer data transfer by communicating read and write requests across the interconnectivity fabric, allowing a device to access and modify another device's local memory without modifying the fabric or its interfaces, using a page table to determine memory locations and translate requests into supported protocols.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If GPUs use system memory for peer-to-peer data transfer, then data can be transferred between GPUs, but transfer speed decreases and overhead increases

Engineering Contradiction:

Improvedata transfer capabilityVSAvoidprocessing speed

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The patent introduces a fabric interface as an intermediary component that enables direct peer-to-peer data transfer between GPUs through the interconnectivity fabric. The fabric interface translates read/write requests into fabric-supported protocols, allowing GPUs to access each other's frame buffers directly without using system memory as an intermediary storage location.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Adaptability or versatility

If existing interconnectivity fabric protocols are used, then system compatibility is maintained, but peer-to-peer reading is not enabled

Engineering Contradiction:

Improveprotocol compatibilityVSAvoiddata transfer efficiency

Core Design Contradiction:

Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent changes the operational parameters of the interconnectivity fabric by introducing a fabric interface that translates standard read/write requests into fabric-supported protocols. This parameter transformation allows the existing fabric infrastructure to support peer-to-peer reading functionality without requiring changes to the fundamental fabric protocol specification.

Inventive Principle:

Principle #35Parameter changes

3Ease of operation

If GPUs duplicate data creation work, then data access is simplified, but processing efficiency decreases

Engineering Contradiction:

Improvedata access simplicityVSAvoidprocessing efficiency

Core Design Contradiction:

Ease of operationVSProductivity

Solution Approach 1:

The fabric interface acts as an intermediary that enables direct access to peer frame buffers, eliminating the need for GPUs to duplicate data creation work. By providing a translation layer that handles peer-to-peer read requests, the system allows GPUs to efficiently share and access each other's processed data without redundant computation.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS7275123B2Method and apparatus for providing peer-to-peer data transfer within a computing environment

Publication Date: 2007.09.25 NVIDIA CORP

AI summary

A method and apparatus for providing peer-to-peer data transfer through an interconnecting fabric. The method and apparatus enable a first device to read and/or write data to/from a local memory of a second device by communicating read and write requests across the interconnectivity fabric. Such data transfer can be performed even when the communication protocol of the interconnectivity fabric does not permit such transfers.