Abstract: Prior knowledge of access pattern is leveraged to improve energy dissipation for general matrix operations. This improves memory access energy for a multitude of applications such as image processing, deep neural networks, and scientific computing workloads, for example. In some embodiments, prior knowledge of access pattern allows for burst read and/or write operations. As such, burst mode solution can provide energy savings in both READ (RD) and WRITE (WR) operations. For machine learning or inference, the weight values are known ahead in time (e.g., inference operation), and so the unused bytes in the cache line are exploited to store a sparsity map that is used for disabling read from either upper or lower half of the cache line, thus saving dynamic capacitance.

Abstract: Prior knowledge of access pattern is leveraged to improve energy dissipation for general matrix operations. This improves memory access energy for a multitude of applications such as image processing, deep neural networks, and scientific computing workloads, for example. In some embodiments, prior knowledge of access pattern allows for burst read and/or write operations. As such, burst mode solution can provide energy savings in both READ (RD) and WRITE (WR) operations. For machine learning or inference, the weight values are known ahead in time (e.g., inference operation), and so the unused bytes in the cache line are exploited to store a sparsity map that is used for disabling read from either upper or lower half of the cache line, thus saving dynamic capacitance.

Abstract: Efficient hardware implementations of a binary search algorithm are provided.

Abstract: Techniques are described for generating asynchronous circuits (e.g., in the form of one or more netlists) for implementation, e.g., in integrated circuitry/chips. Embodiments are directed to asynchronous multi-level domino design template and several variants including a mixture of domino and single-rail data logic. The templates can provide high throughput, low latency, and area efficiency. A multi-level domino template is partitioned into pipeline stages in which each stage consists of potentially multiple-levels of domino logic controlled by a single controller that communicates with other controllers via handshaking. Each stage is composed of two parts: a data path and a control path. The data path implements the computational logic, both combinational and sequential using efficient dual-rail domino logic. The control path implements a unique four-phase handshake to ensure correctness and the preservation of logical dependencies between pipeline stages.

Abstract: Techniques are described for generating asynchronous circuits from any arbitrary HDL representation of a synchronous circuit by automatically clustering the synthesized gates into pipeline stages that are then slack-matched to meet performance goals while minimizing area. Automatic pipelining can be provided in which the throughput of the overall design is not limited to the clock frequency or the level of pipelining in the original RTL specification. The techniques are applicable to many asynchronous design styles. A model and infrastructure can be designed that guides clustering to avoid the introduction of deadlocks and achieve a target circuit performance. Slack matching models can be used to take advantage of fanout optimizations of buffer trees that improve the quality of the results.

Abstract: A memory is described which includes a main memory array made up of multiple single-ported memory banks connected by parallel read and write buses, and a sideband memory equivalent to a single dual-ported memory bank. Control logic and tags state facilitates a pattern of access to the main memory and the sideband memory such that the memory performs like a fully provisioned dual-ported memory capable of reading and writing any two arbitrary addresses on the same cycle.

Abstract: Efficient hardware implementations of a binary search algorithm are provided.

Abstract: Techniques are described for converting netlists for synchronous circuits such as combinational modules, flip flops (or latches), and clock gating modules, to netlist of asynchronous modules. Processes including algorithms are described that bundle multiple modules in an enable domain, so that they are activated only if the incoming enable token to the enable domain has the UPDATE value. The modules can be clustered inside an enable domain, so that each cluster has a separate controller. The objective function of bundling and clustering can minimize power consumption with respect to a given cycle time. Exemplary embodiments can include a gated multilevel domino template.

Abstract: Methods and apparatus are described for optimizing a circuit design. A gate level circuit description corresponding to the circuit design is generated. The gate level circuit description includes a plurality of pipelines across a plurality of levels. Using a linear programming technique, a minimal number of buffers is added to selected ones of the pipelines such that a performance constraint is satisfied.

Abstract: Techniques are described for generating asynchronous circuits (e.g., in the form of one or more netlists) for implementation, e.g., in integrated circuitry/chips. Embodiments are directed to asynchronous multi-level domino design template and several variants including a mixture of domino and single-rail data logic. The templates can provide high throughput, low latency, and area efficiency. A multi-level domino template is partitioned into pipeline stages in which each stage consists of potentially multiple-levels of domino logic controlled by a single controller that communicates with other controllers via handshaking. Each stage is composed of two parts: a data path and a control path. The data path implements the computational logic, both combinational and sequential using efficient dual-rail domino logic. The control path implements a unique four-phase handshake to ensure correctness and the preservation of logical dependencies between pipeline stages.

Abstract: A shared memory is described having a plurality of receive ports and a plurality of transmit ports characterized by a first data rate. A memory includes a plurality of memory banks organized in rows and columns. Operation of the memory array is characterized by a second data rate. Non-blocking receive crossbar circuitry is operable to connect any of the receive ports with any of the memory banks. Non-blocking transmit crossbar circuitry is operable to connect any of the memory banks with any of the transmit ports. Buffering is operable to decouple operation of the receive and transmit ports at the first data rate from operation of the memory array at the second data rate.

Abstract: A shared memory is described having a plurality of receive ports and a plurality of transmit ports characterized by a first data rate. A memory includes a plurality of memory banks organized in rows and columns. Operation of the memory array is characterized by a second data rate. Non-blocking receive crossbar circuitry is operable to connect any of the receive ports with any of the memory banks. Non-blocking transmit crossbar circuitry is operable to connect any of the memory banks with any of the transmit ports. Buffering is operable to decouple operation of the receive and transmit ports at the first data rate from operation of the memory array at the second data rate. Scheduling circuitry is operable to control interaction of the ports, crossbar circuitry, and memory array to effect storage and retrieval of the data segments in the shared memory.

Abstract: A memory is described which includes a main memory array made up of multiple single-ported memory banks connected by parallel read and write buses, and a sideband memory equivalent to a single dual-ported memory bank. Control logic and tags state facilitates a pattern of access to the main memory and the sideband memory such that the memory performs like a fully provisioned dual-ported memory capable of reading and writing any two arbitrary addresses on the same cycle.

Abstract: An asynchronous circuit is described for processing units of data having a program order associated therewith. The circuit includes an N-way-issue resource comprising N parallel pipelines. Each pipeline is operable to transmit a subset of the units of data in a first-in-first-out manner. The asynchronous circuit is operable to sequentially control transmission of the units of data in the pipelines such that the program order is maintained.

Abstract: Techniques are described for generating asynchronous circuits from any arbitrary HDL representation of a synchronous circuit by automatically clustering the synthesized gates into pipeline stages that are then slack-matched to meet performance goals while minimizing area. Automatic pipelining can be provided in which the throughput of the overall design is not limited to the clock frequency or the level of pipelining in the original RTL specification. The techniques are applicable to many asynchronous design styles. A model and infrastructure can be designed that guides clustering to avoid the introduction of deadlocks and achieve a target circuit performance. Slack matching models can be used to take advantage of fanout optimizations of buffer trees that improve the quality of the results.

Abstract: Techniques are described for converting netlists for synchronous circuits such as combinational modules, flip flops (or latches), and clock gating modules, to netlist of asynchronous modules. Processes including algorithms are described that bundle multiple modules in an enable domain, so that they are activated only if the incoming enable token to the enable domain has the UPDATE value. The modules can be clustered inside an enable domain, so that each cluster has a separate controller. The objective function of bundling and clustering can minimize power consumption with respect to a given cycle time. Exemplary embodiments can include a gated multilevel domino template.

Abstract: Methods and apparatus are described for optimizing a circuit design. A gate level circuit description corresponding to the circuit design is generated. The gate level circuit description includes a plurality of pipelines across a plurality of levels. Using a linear programming technique, a minimal number of buffers is added to selected ones of the pipelines such that a performance constraint is satisfied.

Abstract: Methods and apparatus are described for optimizing a circuit design. A gate level circuit description corresponding to the circuit design is generated. The gate level circuit description includes a plurality of pipelines across a plurality of levels. Using a linear programming technique, a minimal number of buffers is added to selected ones of the pipelines such that a performance constraint is satisfied.

Abstract: Methods and apparatus are described relating to a crossbar which is operable to route data from any of a first number of input channels to any of a second number of output channels according to routing control information. Each combination of an input channel and an output channel corresponds to one of a plurality of links. The crossbar circuitry is operable to route the data in a deterministic manner on each of the links thereby preserving a partial ordering represented by the routing control information. Events on different links are uncorrelated.

Abstract: Methods and apparatus are described relating to a crossbar which is operable to route data from any of a first number of input channels to any of a second number of output channels according to routing control information. Each combination of an input channel and an output channel corresponds to one of a plurality of links. The crossbar circuitry is operable to route the data in a deterministic manner on each of the links thereby preserving a partial ordering represented by the routing control information. Events on different links are uncorrelated.