US 12,731,019 B2 · Asynchronous accumulator using logarithmic-based arithmetic

Abstract

Neural networks, in many cases, include convolution layers that are configured to perform many convolution operations that require multiplication and addition operations. Compared with performing multiplication on integer, fixed-point, or floating-point format values, performing multiplication on logarithmic format values is straightforward and energy efficient as the exponents are simply added. However, performing addition on logarithmic format values is more complex. Conventionally, addition is performed by converting the logarithmic format values to integers, computing the sum, and then converting the sum back into the logarithmic format. Instead, logarithmic format values may be added by decomposing the exponents into separate quotient and remainder components, sorting the quotient components based on the remainder components, summing the sorted quotient components using an asynchronous accumulator to produce partial sums, and multiplying the partial sums by the remainder components to produce a sum. The sum may then be converted back into the logarithmic format.

Claims20

Dependent claims

A processing element, comprising:

a set of M vector inference engines, each vector inference engine comprising:
a vector adder configured to compute product values by summing each weight in a set of weights with each activation in a set of activations, wherein the weights and the activations are represented in logarithmic form, each product value including a quotient component and a remainder component;
N arrays of asynchronous accumulators, wherein, within each one of the N arrays:
each asynchronous accumulator corresponds to a different possible value of the remainder component and is configured to increment or decrement a single bit of a partial sum based on the quotient component;
and
a scale unit coupled to each asynchronous accumulator multiplies the partial sum by the value of the remainder component corresponding to the asynchronous accumulator to produce a scaled partial sum;
and
an addition unit adds the scaled partial sums to produce a sum of the product values.

5 more elements

A computer-implemented method for computing a sum of product values using N arrays of asynchronous accumulators, comprising:

receiving a set of weights and a set of activations by a set of M vector adders within M vector inference engines, wherein the weights and the activations are represented in logarithmic form;
summing each weight in the set of weights with each activation in the set of activations by a first one of the M vector adders to compute product values, wherein each product value includes a sign, a quotient component, and a remainder component and each asynchronous accumulator within one of the N arrays corresponds to a different possible value of the remainder component;
either incrementing or decrementing according to the sign, by a first asynchronous accumulator within a first one of the N arrays, a single bit of a partial sum output by the first asynchronous accumulator based on a first one of the quotient components;
multiplying, by a scale unit coupled to the first asynchronous accumulator, the partial sum by the value of the remainder component corresponding to the first asynchronous accumulator to produce a scaled partial sum;
and
adding the scaled partial sum to additional scaled partial sums by an addition unit to produce the sum of the product values.

3 more elements

Dependencies are inferred from the claim text.

Asynchronous accumulator using logarithmic-based arithmetic

Life of this patent

Asynchronous accumulator using logarithmic-based arithmetic

Life of this patent

Abstract

Claims20