Hybrid Floating-Point Format for DSP Silicon Area Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital signal processors and hardware accelerators face inefficiencies in numerical processing due to the complexity of conventional floating-point representations, leading to increased silicon area requirements and power consumption, particularly in wireless communication systems that demand high dynamic range and rapid execution of operations.
Innovation Solution
A hybrid floating-point format is introduced, representing numbers as a sign bit, exponent bits with bias, and mantissa bits, with special handling for zero and overflow conditions, allowing for a variable length format that simplifies exception handling and reduces power consumption while maintaining performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional floating-point representation is used, then numerical processing capability is provided, but silicon area increases and power consumption rises
Solution Approach 1:
The patent extracts and removes unnecessary complexity from conventional floating-point representation by eliminating support for special values (NaN, infinity, denormalized numbers) and using a simplified exponent format without bias encoding. This extraction of essential functionality reduces the silicon area required for implementation while maintaining core numerical processing capability.
Solution Approach 2:
The patent changes key parameters of the floating-point format: using unsigned integers instead of biased integers for exponents, allocating more bits to the mantissa (23 bits vs. conventional 23 bits but with different interpretation), and removing special value encodings. These parameter changes optimize the format for specific DSP applications, reducing hardware complexity and silicon area while maintaining numerical processing effectiveness.
2Productivity
If conventional floating-point representation is used, then numerical operations can be performed, but power consumption increases
Solution Approach 1:
The patent removes complex exception handling logic for special floating-point values (NaN, infinity, denormalized numbers) that present in conventional formats. This extraction eliminates unnecessary computational paths and control logic that would consume power, while retaining full capability for standard numerical operations required in DSP applications.
Solution Approach 2:
By changing the exponent representation to unsigned integers and adjusting the mantissa interpretation, the patent simplifies the arithmetic operations required. This parameter change reduces the complexity of comparison, addition, and multiplication operations, thereby lowering power consumption while maintaining numerical operation capability.
3Area of stationary object
If simplified floating-point format is used, then silicon area is reduced, but handling of special values becomes limited
Solution Approach 1:
The patent adopts a pragmatic approach by designing a floating-point format optimized for typical DSP workloads where special values (NaN, infinity, denormalized numbers) are rarely needed. The simplified format sacrifices support for these edge cases to achieve significant reductions in silicon area and complexity, accepting that the lost versatility has minimal impact on actual DSP performance and energy efficiency.
Data Source
AI summary
An embodiment of a system and method for performing a numerical operation on input data in a hybrid floating-point format includes representing input data as a sign bit, exponent bits, and mantissa bits. The exponent bits are represented as an unsigned integer including an exponent bias, and a signed numerical value of zero is represented as a first reserved combination of the mantissa bits and the exponent bits. Each of all other combinations of the mantissa bits and the exponent bits represents a real finite non-zero number. The mantissa bits are operated on with a “one” bit before a radix point for the all other combinations of the mantissa bits and the exponent bits.