Memory Controller Sparsity Poison for Uncorrectable Error Handling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing error correction schemes for memory systems are inefficient in handling uncorrectable errors (DUEs), leading to productivity loss when processes or kernels are shut down, especially in high-performance computing environments like AI training systems.
Innovation Solution
Convert uncorrectable error data (poison data) into sparsity poison by zeroing out the affected data, allowing compute circuitry to continue operations with reduced accuracy impact, using sparsity poison to maintain productivity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If error correction codes are used to detect errors, then error detection capability is improved, but data bandwidth overhead becomes prohibitively expensive
Solution Approach 1:
The patent extracts the error handling mechanism from the traditional software stack and relocates it to hardware circuitry. The memory controller includes dedicated error detection and handling circuits that operate independently of the main data processing path, allowing error management to be performed without consuming compute resources or shutting down processes.
Solution Approach 2:
The patent introduces an intermediary mechanism (the memory controller with integrated error handling) that sits between the memory and compute circuitry. This intermediary detects and handles errors before they reach the compute stack, preventing process shutdowns while maintaining data integrity through controlled poison data propagation.
2Reliability
If the software stack shuts down the process upon detecting poison data, then data integrity is maintained, but productivity is lost
Solution Approach 1:
The patent applies partial action by allowing poison data to propagate through the system in a controlled manner rather than completely blocking it. The memory controller permits poison data to reach compute circuitry, which then handles it appropriately without shutting down the entire process, achieving partial error containment while maintaining overall system productivity.
Solution Approach 2:
The patent changes the parameter of poison data handling from binary (shutdown/continue) to a spectrum of controlled propagation. By modifying how poison data is treated—allowing it to propagate with controlled impact rather than triggering immediate shutdown—the system maintains both data integrity awareness and process continuity.
3Productivity
If compute operations continue with poison data, then productivity is maintained, but accuracy is degraded
Solution Approach 1:
The patent converts the harmful effect of poison data into a beneficial controlled propagation mechanism. Instead of poison data causing unpredictable corruption, the system allows it to propagate in a controlled manner where its impact can be managed and where it serves as a signal for selective error handling without triggering unnecessary process shutdowns.
Data Source
AI summary
Embodiments herein can avoid shutting down a process that receives poison data that includes an uncorrectable error by converting the poison data into sparsity data. In one embodiment, the sparsity data comprises zeros that replace the bits of the poison data. Compute circuitry can then perform its task as normal, but instead using the zeros of the sparsity data instead of the poison data. Because the poison data is now zeros, they have a reduced negative effect on the process being performed by the compute circuitry.