Floating-Point Memory Scattering for ECC Overhead Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing memory devices face inefficiencies due to the overhead of error correction mechanisms, which occupy additional space without effectively utilizing all available memory locations, and there is a need for techniques to reduce this overhead while maintaining data integrity.
Innovation Solution
A system utilizing a scatter unit to store different portions of floating-point numbers in memory devices with varying error correction capabilities, where more critical fields like the sign and exponent are stored in a highly reliable memory and less critical fields like the mantissa are stored in a less reliable memory, ensuring resilience against errors without significantly impacting calculation results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If high error correction capability is used in memory devices, then data reliability is improved, but memory overhead increases due to more error correction bits
Solution Approach 1:
The patent segments floating-point numbers into different fields (sign, exponent, mantissa) and stores them in different memory devices with different error correction capabilities. Critical fields (sign, exponent) are stored in high-reliability memory while less critical fields (mantissa) are stored in low-reliability memory, reducing overall memory overhead while maintaining data integrity for critical components.
Solution Approach 2:
The patent applies different error correction qualities to different parts of the data structure. High error correction capability is applied locally to critical fields (sign and exponent) that require high reliability, while lower error correction capability is applied to less critical fields (mantissa), optimizing the balance between reliability and memory overhead.
2Reliability
If uniform high reliability storage is used for all data portions, then data integrity is improved, but memory efficiency deteriorates due to excessive error correction overhead
Solution Approach 1:
The patent divides floating-point numbers into separate fields and stores them in different memory locations with different reliability levels. This segmentation allows critical fields to receive high reliability protection while non-critical fields use lower reliability storage, improving memory efficiency without sacrificing data integrity for essential components.
Solution Approach 2:
Different reliability qualities are applied locally to different fields based on their importance. The sign and exponent fields receive high reliability treatment while the mantissa field receives lower reliability treatment, optimizing memory efficiency while maintaining data integrity where it matters most.
3Quantity of substance
If low error correction capability is used in memory devices, then memory overhead is reduced, but data reliability deteriorates
Solution Approach 1:
The patent segments floating-point numbers into critical and non-critical fields, storing critical fields (sign, exponent) in high-reliability memory and non-critical fields (mantissa) in low-reliability memory. This segmentation reduces overall memory overhead while maintaining data reliability for critical components.
Solution Approach 2:
High error correction capability is applied locally to critical fields (sign and exponent) to maintain data reliability, while lower error correction capability is applied to less critical fields (mantissa) to reduce memory overhead, achieving an optimal balance between reliability and overhead.
4Measurement precision
If all floating-point fields are stored in high reliability memory, then calculation accuracy is improved, but memory cost increases
Solution Approach 1:
The patent segments floating-point numbers into different fields and stores them in memory devices with different reliability levels. Critical fields (sign, exponent) are stored in high-reliability memory to maintain calculation accuracy, while less critical fields (mantissa) are stored in low-reliability memory to reduce memory cost.
Solution Approach 2:
High reliability memory is applied locally to critical fields (sign and exponent) that directly impact calculation accuracy, while lower reliability memory is used for less critical fields (mantissa), optimizing the balance between calculation accuracy and memory cost.
Data Source
AI summary
Systems, apparatuses, and methods for scattering floating point values to heterogeneous memory devices are disclosed. An inference engine performs floating point calculations during pre-training and during post-training operations. A scatter unit stores the floating point number values in multiple memories with different error correction capabilities. A first portion of each floating point number value is stored in a first memory having a relatively high error correction capability, and a second portion of each floating point number value is stored in a second memory with a relatively low error correction capability. In one scenario, the first portion includes the sign and exponent fields, while the second portion includes the mantissa field. The resiliency of the inference engine to overcome small errors allows for convergence to the final result in spite of any errors in the retrieved second portion.