Differential Analog MAC Circuit for Signed CiM and PVT Tolerance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Analog mixed-signal compute-in-memory (CiM) processors face challenges in supporting signed multi-bit data and have low process, voltage, and temperature (PVT) variation tolerance, leading to increased memory requirements and reduced computing precision, especially in edge devices with power limitations.
Innovation Solution
The implementation of a differential signal path with butterfly switch circuitry and two-rail capacitor ladder networks enables signed multiply-accumulate operations, providing intrinsic PVT variation tolerance and reducing memory requirements by representing signed data without doubling memory cells, and using differential signaling to cancel noise and variations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If analog mixed-signal hardware is used for MAC operations, then computing efficiency is improved, but PVT variation tolerance deteriorates
Solution Approach 1:
The patent segments the analog computation process into differential signal paths that separately process positive and negative weight values. By dividing the computation into two complementary signal paths (one for positive weights, one for negative weights), the system achieves both high computing efficiency and robustness against PVT variations through differential signaling that cancels common-mode noise and variations.
Solution Approach 2:
The patent changes the parameter representation by using differential voltage signals instead of single-ended analog voltages. This parameter transformation from single-ended to differential signaling fundamentally improves PVT tolerance while maintaining analog computing efficiency, as the differential nature cancels out process, voltage, and temperature variations that affect both signals equally.
2Adaptability or versatility
If signed multi-bit data is supported, then data representation capability is improved, but memory requirements increase
Solution Approach 1:
The patent adds a dimensional transformation by mapping signed multi-bit data into the differential signal domain. Instead of requiring separate memory cells for each bit (including sign bit), the invention uses the differential voltage dimension to encode signed values, where the voltage difference between two rails represents the signed magnitude, effectively using a voltage dimension to store what would otherwise require additional memory bits.
Solution Approach 2:
The patent uses a copying approach where unsigned weight values are copied into the differential signal path twice - once for positive weights and once for negative weights. This copying strategy allows the same memory resources to be reused for both positive and negative representations, eliminating the need for separate memory cells for signed data while maintaining full signed multi-bit data representation capability.
3Measurement precision
If calibration units are added to improve precision, then computing precision is improved, but device complexity and power consumption increase
Solution Approach 1:
The patent implements self-service by designing the differential signal path to inherently cancel PVT variations and noise through its differential nature. The system does not require external calibration units or correction circuits because the differential architecture automatically compensates for errors, making the computation precision self-maintaining without additional complex calibration hardware.
Solution Approach 2:
The patent converts the harmful effect of PVT variations and noise into a beneficial feature by exploiting the differential signaling property where common-mode variations appear equally on both rails and are naturally rejected during subtraction. What would normally be harmful noise and variations become a mechanism for automatic error cancellation, eliminating the need for separate calibration units.
Data Source
AI summary
Systems, apparatuses and methods may provide for technology that conducts, by a differential signal path, signed multiply-accumulate (MAC) operations on first analog signals and multibit weight data stored in the differential signal path, and outputs, by the differential signal path, second analog signals based on the signed MAC operations.