Multi-Format MAC Alignment for Low-Latency Floating-Point Accumulation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems face long processing times and data hazards in high-speed operations involving floating point data, particularly in mixed precision DNN accelerators, due to the need for lengthy floating point addition processes.
Innovation Solution
An apparatus and method that supports multi-format data processing by using a receiver, processors with multipliers, and aligners to perform alignments based on exponent values, enabling efficient multiplication and addition operations with reduced latency and data hazards, utilizing a multiplier-accumulator array for unsigned multiplications and custom formatting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If floating point adder is used for accumulation, then accumulation operation can be performed, but processing time becomes long causing data hazard in pipeline
Solution Approach 1:
The accumulation operation is segmented into multiple stages: first alignment based on exponent values, addition, and second alignment based on exponent value and previous cycle result. This segmentation allows each stage to be optimized independently and enables pipeline execution, eliminating data hazards while maintaining processing efficiency
Solution Approach 2:
The first alignment is performed preliminarily before addition based on exponent values, and the second alignment is performed after addition based on previous cycle results. These preliminary actions prepare data in advance for subsequent operations, enabling overlapping execution and eliminating pipeline stalls
2Adaptability or versatility
If multi-format data support is implemented, then versatility is improved, but device complexity increases
Solution Approach 1:
The apparatus is designed with universal components that can handle multiple data formats through a unified architecture. The receiver accepts various data formats, and the processors perform multiplication, alignment, and addition operations that work across different formats using the same core components, avoiding the need for separate dedicated circuits for each format
Solution Approach 2:
The system manages different data formats by changing parameters such as exponent values and alignment shifts rather than requiring different hardware architectures. The same hardware components adapt to handle various formats by adjusting control parameters, reducing structural complexity while maintaining versatility
Data Source
Figure 1A
Figure 1B
Figure 2
AI summary
An apparatus with multi-format data support includes: a receiver configured to receive a plurality of data corresponding to a plurality of data formats; one or more processors configured to: multiply the plurality of data using one or more multipliers; perform a first alignment on a result of the multiplication based on an exponent value of the plurality of data; add a result of the first alignment; and perform a second alignment on a result of the addition based on the exponent value and an operation result of a previous cycle.