摘要:
Methods, apparatus, systems, and articles of manufacture are disclosed for high throughput compression of neural network weights. An example apparatus includes at least one memory, instructions in the apparatus and processor circuitry to execute the instructions to determine sizes of data lanes in a partition of neural network weights, determine a slice size based on a size difference between a first data lane and a second data lane of the data lanes in the partition, the first data lane including first data, the second data lane including second data, the second data of a smaller size than the first data, cut a portion of the first data from the first data lane based on the slice size, and append the portion of the first data to the second data lane.
摘要:
Described is an apparatus that comprises: a first sequential unit; a first queue coupled in parallel to the first sequential unit such that the first queue and first sequential unit receive a first input, the first sequential for double sampling the first input; a compare unit to receive an output from the first sequential unit; and a first selection unit controllable by a write pointer of a previous cycle, the first selection unit to receive outputs of each storage unit of the first queue, wherein the first selection unit to generate an output for comparison by the first compare unit.
摘要:
Systems and methods for performing encoding and/or decoding can include an input data path that receives multiple input data values having an order (significance) with respect to one another. Each input data value can be applied to multiple compute paths (106-1 to 106-N), each of which can precompute multiple output values based on a different predetermined disparity value. Multiplexers (114-1 to 114-N) can output one precomputed output value according to a disparity value corresponding to a previous input data value in the order.
摘要:
A circuit configured to provide a storage device comprising one or more virtual multiqueue FIFOs. The circuit is generally configured to operate at a preferred clock speed of a plurality of clock speeds.
摘要:
Technologies for memory management of a neural network include a compute device to read a memory of the compute device to access connectivity data associated with a neuron of the neural network, determine a memory address at which weights corresponding with the one or more network connections are stored, and access the corresponding weights from a memory location corresponding with the memory address. The connectivity data is indicative of one or more network connections from the neuron.
摘要:
An apparatus is described. The apparatus includes a semiconductor chip. The semiconductor chip includes spiking neural network circuitry. The spiking neural network circuitry includes a learning rule circuit. The learning rule circuit includes a resistive element. A resistance of the resistive element is to determine a change in a weight of a synapse between neurons of the spiking neural network circuitry.
摘要:
A packet processing system including an encapsulator engine, and a packet pre-processor coupled to the encapsulator engine. The packet pre-processor calculates a variation between an input data rate and a pre-determined output data rate. The input data rate is based on a number of data read requests. The packet pre-processor compensates for the variation by modifying the number of data read requests.
摘要:
Disclosed is an apparatus and method for an interface for scheduling the store-forwarding of back-to-back multi-channel packet fragments. The apparatus may include a data extraction block, a plurality of data assemblers, a scheduler, and an output memory. The scheduler may be configured to operate according to a scheduling policy. The scheduling policy may include a set of priorities to determine an order of scheduling writes to the output memory from a plurality of data assemblers. The scheduling policy may also include selecting one or more of the plurality of data assemblers having a fill level greater than twice an input data path width of an input data bus and having no end-of-packet (EOP) or start-of-packet (SOP) as a first priority, selecting one or more of the plurality of data assemblers having a fill level greater than twice the input data path width of the input data bus, and not covered in the first priority selection, as a second priority, selecting one or more of the plurality of data assemblers having a fill level greater than the input data path width of the input data bus, and not covered in the first and second priority selections, as a third priority, and selecting one or more of the plurality of data assemblers having an end-of-packet (EOP) as a fourth priority.
摘要:
Disclosed is an apparatus and method used in an interface for scheduling the store-forwarding of back-to-back multi-channel packet fragments, comprising a first-in-first-out (FIFO) memory, a read pointer of the FIFO memory, the read pointer to increment by at least one of a unit access per read and a fraction of the unit access per read, and a write pointer of the FIFO memory, the write pointer to increment by at least one of a unit access per write and a fraction of the unit access per write.
摘要:
In one embodiment, the invention is an apparatus. The apparatus includes a FIFO array having a first plurality of memory elements, each memory element having a predetermined number of bits, the FIFO array having a read pointer. The apparatus also includes a FIFO control register array having a second plurality of memory elements, each memory element of the second plurality corresponding to a memory element of the first plurality of memory elements, the read pointer suitable for accessing the FIFO control register array. The apparatus further includes a control logic block coupled to the FIFO control register array and the FIFO array. The control logic block is to receive a data value of the memory element of the FIFO control register array pointed to by the read pointer. The control logic block is also to signal the read pointer to stall responsive to the data value having a first value.