Abstract: Embodiments provide mechanisms to facilitate compute operations for deep neural networks. One embodiment comprises a graphics processing unit comprising one or more multiprocessors, at least one of the one or more multiprocessors including a register file to store a plurality of different types of operands and a plurality of processing cores. The plurality of processing cores includes a first set of processing cores of a first type and a second set of processing cores of a second type. The first set of processing cores are associated with a first memory channel and the second set of processing cores are associated with a second memory channel.

Abstract: A carry lookahead adder having a reduced internal block fanout which is achieved efficiently in terms of the silicon area needed to implement the carry lookahead adder. The carry lookahead adder of the present invention is characterized by a modified tree structure having carry generate/propagate signal operators located in such a manner that the maximum internal block fanout is equal to (adder width)/8 for adders having a width of at least 16 bits. For adders having a width of less than 16 bits, the internal block fanout is 2. The routing complexity is increased in order to implement redundant overlapping carry generate/propagate operations which, in turn, decreases the internal block fanout of the adder. However, increases in routing complexity can be accomplished within the minimum X-by-Y area of each stage of the adder. Therefore, the overall performance of the carry lookahead adder of the present invention can be optimized while meeting minimum area requirements.

Abstract: An incrementer/decrementer architecture having a reduced internal block fanout which is achieved efficiently in terms of the silicon area needed to implement the incrementer/decrementer. The incrementer/decrementer of the present invention is characterized by a modified tree structure having operators located in such a manner that the maximum internal block fanout is equal to (incrementer/decrementer width)/8 for incrementer/decrementers having a width of at least 16 bits. For incrementer/decrementers having a width of less than 16 bits, the internal block fanout is 2. The routing complexity is increased in order to implement redundant overlapping operations which, in turn, decreases the internal block fanout. However, increases in routing complexity can be accomplished within the minimum X-by-Y area of each stage of the incrementer/decrementer. Therefore, the overall performance of the incrementer/decrementer of the present invention can be optimized while meeting minimum area requirements.