Abstract: A plurality of chiplets may be used to multiply two matrices A and B. Matrix A may be decomposed into horizontal stripes and matrix B may be decomposed into vertical stripes. Each of the horizontal stripes may be multiplied by each of the vertical stripes to form the output matrix C. Specifically, horizontal stripes may be stored in a stationary, distributed manner across the chiplets, while the vertical stripes (or sub-vertical stripes) may be passed between respective pairs of the chiplets until each of the vertical stripes (or sub-vertical stripes) of matrix B has been received and processed by each of the chiplets. The vertical stripes may be passed along one or more paths that interconnect the chiplets. Similar techniques can be applied to an arrangement in which the vertical stripes are stationary and the horizontal stripes (or sub-horizontal stripes) are passed between respective pairs of the chiplets.

Abstract: Dynamic data quantization may be applied to minimize the power consumption of a system that implements a convolutional neural network (CNN). Under such a quantization scheme, a quantized representation of a 3×3 array of m-bit activation values may include 9 n-bit mantissa values and one exponent shared between the n-bit mantissa values (n<m); and a quantized representation of a 3×3 kernel with p-bit parameter values may include 9 q-bit mantissa values and one exponent shared between the q-bit mantissa values (q<p). Convolution of the kernel with the activation data may include computing a dot product of the 9 n-bit mantissa values with the 9 q-bit mantissa values, and summing the shared exponents. In a CNN with multiple kernels, multiple computing units (each corresponding to one of the kernels) may receive the quantized representation of the 3×3 array of m-bit activation values from the same quantization-alignment module.

Abstract: Dynamic data quantization may be applied to minimize the power consumption of a system that implements a convolutional neural network (CNN). Under such a quantization scheme, a quantized representation of a 3×3 array of m-bit activation values may include 9 n-bit mantissa values and one exponent shared between the n-bit mantissa values (n<m); and a quantized representation of a 3×3 kernel with p-bit parameter values may include 9 q-bit mantissa values and one exponent shared between the q-bit mantissa values (q<p). Convolution of the kernel with the activation data may include computing a dot product of the 9 n-bit mantissa values with the 9 q-bit mantissa values, and summing the shared exponents. In a CNN with multiple kernels, multiple computing units (each corresponding to one of the kernels) may receive the quantized representation of the 3×3 array of m-bit activation values from the same quantization-alignment module.

Abstract: Dynamic data quantization may be applied to minimize the power consumption of a system that implements a convolutional neural network (CNN). Under such a quantization scheme, a quantized representation of a 3×3 array of m-bit activation values may include 9 n-bit mantissa values and one exponent shared between the n-bit mantissa values (n<m); and a quantized representation of a 3×3 kernel with p-bit parameter values may include 9 q-bit mantissa values and one exponent shared between the q-bit mantissa values (q<p). Convolution of the kernel with the activation data may include computing a dot product of the 9 n-bit mantissa values with the 9 q-bit mantissa values, and summing the shared exponents. In a CNN with multiple kernels, multiple computing units (each corresponding to one of the kernels) may receive the quantized representation of the 3×3 array of m-bit activation values from the same quantization-alignment module.