Compute-in-Memory Zero Skipping for Sparse Operand Power Savings
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional compute-in-memory circuits are inefficient due to energy consumption during computations where one or more operands are zero, as they do not compensate for zero input data, leading to unnecessary power consumption.
Innovation Solution
Implementing a zero-detection circuit in CIM circuits to detect when at least one operand is zero, disabling components such as memory access, multiplication, and mantissa shift operations, thereby reducing power consumption while preserving computational accuracy and throughput.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If conventional compute-in-memory circuits perform computations without zero detection, then computational throughput is maintained, but power consumption increases due to unnecessary operations on zero operands
Solution Approach 1:
The zero-detection circuit performs preliminary detection of zero operands before the main computation operations (memory access, multiplication, mantissa shift) are executed. By detecting zeros in advance, the system can disable subsequent energy-consuming operations, thereby reducing power consumption without significantly impacting computational throughput since the detection overhead is minimal compared to the full computation pipeline.
2Use of energy by moving object
If zero-detection circuit is implemented to disable components, then power consumption is reduced, but device complexity increases
Solution Approach 1:
The zero-detection circuit acts as an intermediary component between the input data and the main computation units (memory array, multiplier, mantissa shift circuit). This intermediary detects zero conditions and generates control signals that enable or disable the downstream computation components, thereby reducing power consumption while adding only a single-layer overhead to the existing circuit architecture.
3Loss of energy
If conventional circuits process all operands uniformly, then computational accuracy is maintained, but energy efficiency deteriorates due to processing zero operands
Solution Approach 1:
The system applies different processing quality to different operands by detecting zero values and applying a specialized processing path (disabling computation) only to those specific cases. Non-zero operands continue to undergo full computation operations. This local differentiation maintains computational accuracy for all operands while significantly improving energy efficiency by avoiding unnecessary operations on zero operands.
Data Source
AI summary
A system, circuit, and method of operation of the system and circuit are disclosed. In one aspect, a device includes a computation circuit, a memory array, and a controller. The controller can determine that one or more input data bits to the computation circuit or one or more memory bits provided from the memory array are all in a first logic state. In response to determining that the one or more input data bits or the one or more memory bits are all in the first logic state, the controller can generate a control signal to disable at least one component of the computation circuit.