Virtual GPC Abstraction for Flexible GPU Resource Migration

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing GPU chip designs face low manufacturing yield due to defects, leading to inefficient use of functional circuitry and complications in software migration across chips with varying TPC per GPC profiles, especially in cloud-based environments.

Innovation Solution

Implementing Virtual GPCs, Flexible TPC and GPC Migration, and Dynamic TPC Disablement to allow chips with differing TPC/GPC profiles to be included in the same product SKU, ensuring consistent programmer view and enabling seamless software migration without full resets.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Power

If chips with more TPCs per GPC are manufactured to increase performance, then processing capability is improved, but manufacturing yield decreases due to higher probability of defects

Engineering Contradiction:

Improveprocessing capabilityVSAvoidmanufacturing yield

Core Design Contradiction:

PowerVSReliability

Solution Approach 1:

The patent divides the GPU into multiple independent GPCs, each containing multiple TPCs. This segmentation allows individual GPCs or TPCs to be disabled if defects are detected, while other segments remain functional. The compute work distributor can selectively activate only the healthy segments, resolving the contradiction between having more processing units (which reduces yield) and maintaining functionality (which requires disabling defective units).

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces dynamic configuration parameters that allow the number of active GPCs and TPCs to be changed at runtime. By modifying the operational parameters of the GPU architecture, the system can adapt to different defect scenarios and maximize the use of functional hardware, thereby improving yield while maintaining performance through parameter-based flexibility rather than fixed manufacturing specifications.

Inventive Principle:

Principle #35Parameter changes

2Reliability

If floorsweeping is applied to disable defective circuitry and maintain consistency across chips, then manufacturing yield is improved, but the number of functional circuitry is reduced

Engineering Contradiction:

Improvemanufacturing yieldVSAvoidfunctional circuitry

Core Design Contradiction:

ReliabilityVSPower

Solution Approach 1:

The patent implements dynamic enablement and disablement of GPCs and TPCs based on actual defect detection. Rather than applying static floorsweeping that permanently disables circuitry, the system dynamically configures which segments are active based on the specific defects present in each chip. This allows maximum functional circuitry to be utilized while maintaining consistency across chips with different defect profiles.

Inventive Principle:

Principle #15Dynamics

3Adaptability or versatility

If dynamic reconfiguration of GPU resources is implemented to accommodate varying TPC/GPC profiles, then adaptability is improved, but device complexity increases

Engineering Contradiction:

Improveflexibility in configurationVSAvoidreconfiguration mechanism

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The compute work distributor automatically performs the complex task of determining which GPCs and TPCs should be active based on defect information, without requiring manual intervention or complex external reconfiguration mechanisms. The system self-configures by interpreting defect maps and automatically adjusting its internal architecture, thereby achieving high adaptability while minimizing the complexity burden on external systems.

Inventive Principle:

Principle #25Self-service

4Ease of operation

If flexible migration between GPU processing clusters is enabled without full resets, then ease of operation is improved, but reliability challenges arise in maintaining consistent profiles

Engineering Contradiction:

Improvesoftware migration capabilityVSAvoidprofile consistency

Core Design Contradiction:

Ease of operationVSReliability

Solution Approach 1:

The patent introduces virtual GPCs as an intermediary layer between the physical hardware and the software runtime. Virtual GPCs provide a standardized, consistent interface that abstracts away the underlying physical variations and defect configurations. This intermediary enables seamless migration of software contexts between different physical GPU configurations while maintaining reliable and consistent profiles, as the virtual layer translates between diverse physical realities and a uniform software abstraction.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS12449473B2Virtualizing hardware processing resources in a processor

Publication Date: 2025.10.21 NVIDIA CORP

AI summary

Processing hardware of a processor is virtualized to provide a façade between a consistent programming interface and specific hardware instances. Hardware processor components can be permanently or temporarily disabled when not needed to support the consistent programming interface and/or to balance hardware processing across a hardware arrangement such as an integrated circuit. Executing software can be migrated from one hardware arrangement to another without need to reset the hardware.