Noise and bound management for RPU array
A method, computer program product, and circuit are provided for noise and bound management for a Resistive Processing Unit (RN) array having an op-amp. The method includes reducing the noise in an output signal from the RPU array by using a largest value, in a sigma vector having a plurality of values, as a representation for a window for an input signal to the RPU array. The input signal to the RPU array is formed from the plurality of values. The method further includes sensing saturation at an output of the op-amp. The method also includes managing the bound to eliminate the saturation by reducing the plurality of values from which the input sign to the RPU is formed.
1. A method for noise and bound management for a Resistive Processing Unit (RPU) array having an op-amp, the method comprising:
reducing the noise in an output signal from the RPU array by using a largest value, in a sigma vector having a plurality of values, as a representation for a window for an input signal to the RPU array, the input signal to the RPU array being formed from the plurality of values;
sensing saturation at an output of the op-amp; and
managing the bound to eliminate the saturation by reducing the plurality of values from which the input signal to the RPU is formed.
2. The method of claim 1 , wherein the RPU array is configured to perform an analog vector-matrix multiplication.
3. The method of claim 1 , wherein the reducing step further comprises scaling remaining ones of the plurality of values with respect to the largest value.
4. The method of claim 1 , wherein the managing step reduces the plurality of values in one or more additional forward passes by the RPU array.
5. The method of claim 1 , wherein the reducing step further comprises:
dividing the sigma vector by the largest value to obtain a quotient and using the quotient as the input signal to the RPU array; and
multiplying the output signal from the RPU array by the largest value to obtain a product.
6. The method of claim 5 , wherein the quotient forms a conditioned input signal to the RPU array, and the product forms a conditioned output signal from the RPU array.
7. The method of claim 5 , wherein the RPU array is configured to perform an analog vector-matrix multiplication using the quotient as an input signal and the product as the output signal.
8. The method of claim 1 , wherein the sigma vector is comprised in a digital signal having a plurality of digital values, and the input to the RPU unit comprises a plurality of analog signal pulse width values each of which represents a respective one of the plurality of digital values.
9. The method of claim 1 , wherein the managing step is performed iteratively until a number of clipped outputs from the op-amp is below a threshold.