Multi-Precision Computation via Mantissa Arrangement Generator
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional GPUs are inefficient in performing double-precision computations, as they are typically configured for single-precision operations, limiting their ability to handle general compute operations effectively.
Innovation Solution
The system employs a plurality of single precision multiply-add units (MADDs) and an arrangement generator to split double precision numbers into mantissa arrangements, enabling efficient double precision computations by loading these arrangements into the MADDs to perform double precision operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single operating block specially designed for double-precision computations is used, then double-precision computation capability is improved, but the device complexity increases and efficiency for graphics data operations deteriorates
Solution Approach 1:
The patent makes existing single-precision MADD units perform double-precision operations by reconfiguring how data is arranged and fed to them. The same hardware blocks handle both single-precision graphics operations and double-precision compute operations, eliminating the need for separate dedicated double-precision operating blocks while maintaining full functionality for both operation types.
Solution Approach 2:
The patent divides double-precision numbers into multiple single-precision components (splitting 64-bit double precision into two 32-bit single precision values). This segmentation allows the existing single-precision MADD units to process double-precision data by handling the split components separately and combining results, thereby avoiding the need for complex dedicated double-precision hardware blocks.
2Device complexity
If conventional single precision MADDs are used for double precision computations, then device complexity is reduced, but computation accuracy deteriorates
Solution Approach 1:
The patent performs preliminary arrangement of double-precision data into specific mantissa configurations before feeding them to the MADD units. The arrangement generator pre-processes the input data, organizing it in a format that enables accurate double-precision computation through subsequent single-precision operations, thereby maintaining computation accuracy without requiring dedicated double-precision hardware.
Solution Approach 2:
The arrangement generator acts as an intermediary component that transforms double-precision input data into a format suitable for single-precision MADD processing. It mediates between the double-precision computational requirements and the single-precision hardware capabilities, ensuring accurate results are achieved through proper data arrangement and multiple computational steps.
3Ease of operation
If single precision operations are used in GPUs, then ease of operation is improved, but double precision computation capability deteriorates
Solution Approach 1:
The patent dynamically reconfigures the operation mode of existing MADD units based on the computational requirements. The same hardware blocks can operate in single-precision mode for graphics data or be reconfigured for double-precision operations through the arrangement generator, providing adaptability without sacrificing the ease of single-precision operation efficiency.
Data Source
AI summary
Systems and methods for multi-precision computation are disclosed. One embodiment of the present invention includes a plurality of multiply-add units (MADDs) configured to perform one or more single precision operations and an arrangement generator to generate one or more mantissa arrangements using a plurality of double precision numbers. Each MADD is configured to receive and load said mantissa arrangements from the arrangement generator. The MADDs compute a result of a multi-precision computation using the mantissa arrangements. In an embodiment, the MADDs are configured to simultaneously perform operations that include, single precision operations, double-precision additions and double-precision multiply and additions.