Abstract: The present disclosure describes an in-memory computing (IMC) circuit including a first set of IMC cells having a first number of IMC cells and a second set of IMC cells having the first number of IMC cells. The first set of IMC cells can generate a first bit-product of a weight number having a first number of bits and a first bit of an input number having a second number of bits. The second set of IMC cells can generate a second bit-product of the weight number and a second bit of the input number. A first IMC cell of the first set of IMC cells includes a first bit-wise multiplication circuit configured to multiply a first bit of the weight number and the first bit of the input number.

Abstract: Embodiments relate to performing multiply-accumulator operation on asymmetrically quantized input data and kernel data in a neural processor. Instead of adjusting to the input data at a multiply-accumulator to account for the asymmetric quantization of the input data, an adjusted bias for the multiply-accumulator operation is computed beforehand and stored in the multiply-accumulator. On the other hand, kernel coefficients derived from the kernel data are adjusted at the multiply-accumulator to account for the asymmetric quantization. In this way, computational complexity associated with asymmetric quantization may be reduced while increasing the efficiency of the convolution operations at the neural processor.

Abstract: Embodiments of the present disclosure relate to a multiply-accumulator circuit that includes a main multiplier circuit operable in a floating-point mode or an integer mode and a supplemental multiplier circuit that operates in the integer mode. The main multiplier circuit generates a multiplied output that undergoes subsequent operations including a shifting operation in the floating-point mode whereas the supplemental multiplier generates another multiplied output that does not undergo any shifting operations. Hence, in the integer mode, two parallel multiply-add operations may be performed by the two multiplier circuits, and therefore accelerate the multiply-adder operations. Due to the lack of additional shifters associated with the supplemental multiplier circuit, the multiply-accumulator circuit does not have a significantly increased footprint.

Abstract: An integrated circuit can include absolute difference circuitry configured to compute an absolute difference value. The absolute difference circuitry may include a comparison logic, a single adder, and a multiplexer. The comparison logic may receive a first input value and a second input value and may generate a comparison value based on whether the first input value exceeds the second input value. The adder may compute a sum of the first input value, an inverted version of the second input value, and the comparison value. The multiplexer may receive the sum and an inverted version of the sum and may output either the sum or the inverted version of the sum based on the comparison value to produce the absolute difference value.

Abstract: Techniques are disclosed relating to asymmetric circuits. In some embodiments, a storage element is configured to maintain a first input value as an input to an asymmetric circuit during a time interval. For example, in one embodiment, the time interval may correspond to a frame of video data and the storage element may be configured to store a filter coefficient for the frame of video data. In some embodiments, the storage element may be configured to store the value as a constant for multiple operations by the asymmetric circuit. In some embodiments, the asymmetric circuit is configured to generate a plurality of output values based on the first input value and respective ones of a set of second input values. In some embodiments, the asymmetric circuit is leakage power asymmetric and/or critical path asymmetric. This may increase performance and/or reduce power consumption.

Abstract: Techniques are disclosed relating to asymmetric circuits. In some embodiments, a storage element is configured to maintain a first input value as an input to an asymmetric circuit during a time interval. For example, in one embodiment, the time interval may correspond to a frame of video data and the storage element may be configured to store a filter coefficient for the frame of video data. In some embodiments, the storage element may be configured to store the value as a constant for multiple operations by the asymmetric circuit. In some embodiments, the asymmetric circuit is configured to generate a plurality of output values based on the first input value and respective ones of a set of second input values. In some embodiments, the asymmetric circuit is leakage power asymmetric and/or critical path asymmetric. This may increase performance and/or reduce power consumption.