In-Memory Computation Drift Compensation for Accurate MACs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In-memory computation devices suffer from computational accuracy issues due to time drifts and temperature variations affecting the computational weights stored in memory cells, leading to errors in current measurements.
Innovation Solution
An in-memory computation device with a word line activation circuit, biasing circuit, memory array, and digital detector that compensates for drifts by adjusting the bias voltage based on a reference current, using a memory array with phase change material cells and a digital detector to sample and convert the bit line current, maintaining accurate output signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If in-memory computation devices use memory cells to store computational weights and perform MAC operations, then data transfer bandwidth is improved and power consumption is reduced, but computational accuracy deteriorates due to time drifts of computational values
Solution Approach 1:
The patent applies preliminary action by performing drift compensation measurements during an idle period before the actual computation. The system measures the drift of computational values stored in memory cells during a time window when no computation is performed, then uses this measured drift to compensate for accuracy errors in subsequent MAC operations. This preliminary measurement and compensation approach maintains high computational accuracy while preserving the bandwidth and power advantages of in-memory computation.
2Speed
If in-memory computation devices perform MAC operations with stored computational weights, then processing speed is improved, but measurement precision deteriorates due to errors in current measurements from time drifts
Solution Approach 1:
The patent implements feedback by continuously monitoring the drift of computational values in memory cells and using this information to adjust and compensate for measurement errors. The system measures drift during idle periods, feeds this drift information back into the computation process, and applies compensation to correct current measurement errors during MAC operations. This feedback mechanism maintains high processing speed while improving current measurement accuracy through dynamic error correction.
Data Source
AI summary
An in-memory computation device includes a word line activation circuit that receives an input signal indicative of input values and provides activation signals each as a function of the input value. The in-memory computation device further includes a memory array, a biasing circuit generating a bias voltage and a digital detector. The memory array has memory cells coupled to a bit line and each to a word line. Each memory cell stores a computational weight. In response to an activation signal, a cell current flows through each memory cell as a function of the bias voltage, the activation signal and the computational weight. A bit line current flows through the bit line as a function of a summation of the cell currents. The digital detector is coupled to the bit line, samples the bit line current and, in response, provides an output signal.